> ## Content Index > Fetch the complete content index at: https://adikumar.co/llms.txt > Use this file to discover other available public pages before exploring further. # The Capital Stack of AI Infrastructure - URL: https://adikumar.co/capital-stack-ai-infrastructure-2026-28/ - Published: 2026-09-01T09:03:41.000Z - Updated: 2026-09-01T14:15:57.000Z - Description: Bottom-up build of the five-layer, five-source capital stack financing AI infrastructure 2026-28. Three headline measurements ($0.65T financing · $0.95T formation+procurement · $0.16T refi) plus the framework for a quarterly Ledger. - Author: Adi Kumar - Tags: Financing, AI infrastructure, Capital markets, Nvidia, Deal Watch, private markets, Datacenter AI Infrastructure Financing · Volume 1 of 6 · Anchor essay for the recurring quarterly Ledger --- [ 5 min Executive brief TL;DR + warning signs + falsifiers ](#five-minute-version) [ 20 min Core analysis Three headline numbers, sources, transmission, duration ](#core-analysis) [ 72 min Full model Bottom-up construction, audit table, stress cascade, Ledger spec ](#full-model) ## The 5-minute version - The AI infrastructure financing stack is becoming legally modular faster than it is becoming economically independent. Separated claims exist across grid, sites, equipment, compute cash flows, and operator equity. Whether that separation constitutes actual risk transfer is not yet demonstrated. - Third-party structures finance roughly $0.65 trillion of gross activity supporting roughly $0.95 trillion of formation and procurement over 2026-28\. Hyperscaler self-funded capex adds another $2.5-3.2 trillion over the same window. - Four principal cross-layer capital providers plus NVIDIA as a strategic-contingent participant underwrite positions across the stack. The concentration lives in the capital-provider set even where the underlying assets are independent. - The 2028-29 refinancing wave is the systemic-risk point. Roughly 2.5-year debt on 5-year GPU collateral against 3-7 year offtake requires the market to reprice mid-cycle against aging assets and expiring contracts. - Underwriters should be tracking how much economic downside actually leaves the hyperscaler balance sheet when demand softens. The Ledger reports this quarterly via the Risk Transfer Ratio (H.6) and Recourse-Adjusted Financing (H.7). ⚠ Six warning signs - **Legal separation is not risk transfer.** Take-or-pay commitments, minimum-purchase guarantees, and cross-recourse keep economic downside on hyperscaler balance sheets even where the paper documents apparent separation. F6 is the falsifier the whole thesis depends on and it is not yet observable. - **The 2028-29 refi cliff.** 2.5-year debt matures against GPU collateral roughly one generation behind and offtake with 6-24 months of remaining tenor. The whole architecture depends on the next lender showing up. If they do not, the cascade is fast (§13). - **Four capital providers hold most of the stack.** Blackstone, MGX, KKR, and Brookfield span three or more stack categories each; NVIDIA adds a fifth via the 6-sponsor platform. A 20% pullback by any one of them hits multiple layers in the same quarter. - **NVIDIA is financing its own demand.** Revenue + ecosystem-financing + direct strategic-support commitments all capitalise the same AI-demand assumption; a demand-side shock reaches all three vectors together (§9). - **Compute (L4) is 46% of the aggregate number and rests on two Grade-E inputs.** Non-hyperscaler share (33% central) and hyperscaler-adjacent overlap (29% central) jointly determine $437B. A 10-point shift in either moves L4 by $60-90B. - **48% of formation is not identified in the Sources taxonomy.** The gap is real, unexplained by any single mechanism, and grows as internal cash flows and hyperscaler-adjacent structures scale outside public disclosure. What would falsify this thesis - **F1 · Third-party financing stays immaterial.** External New-Money Coverage Ratio remains structurally below \~20% through 2028 despite continued formation. The buildout is still hyperscaler-dominated, not multi-layer. - **F2 · Claims stay bundled.** GPU, site, offtake, and operator risk continue to be co-underwritten on the same balance sheet rather than independently priced. - **F3 · GPU collateral never becomes financeable without hyperscaler guarantees.** The Blue Owl-IREN print proves unrepresentative. - **F4 · Cross-layer capital-provider concentration reverses.** Top-four share falls below \~40% or new independent lenders materially broaden the provider set. - **F5 · The 2028-29 refinancing window clears without stress.** First-generation debt refinances at stable or tighter spreads despite aging collateral. - **F6 · Risk transfer does not occur despite financing separation.** Coverage Ratio rises but the Risk Transfer Ratio (H.6) and Recourse-Adjusted Financing (H.7) stay low because hyperscalers keep bearing the economic downside. This is the falsifier the whole architecture depends on. Framework reference · used throughout this essay FIVE PHYSICAL LAYERS (uses side) - **L1a** Grid access · interconnection deposits, transmission upgrades - **L1b** Firm supply · SMR, advanced-reactor, BTM gas - **L2** Site + shell + civil · land, foundations, building shell - **L3** Equipment · transformers, switchgear, UPS, cooling, PDUs - **L4** Compute & IT procurement · GPU + adjacent IT FIVE FINANCING SOURCES - **S1** Customer prepayments · offtake advances from model cos / hyperscalers - **S2** Vendor finance · equipment credit, extended payment terms - **S3** GPU-collateralised debt · senior debt against GPU fleets + offtake - **S4** RE / project-finance debt · site + BTM generation debt - **S5** Sponsor equity · growth equity + sovereign co-invest + listed operator **Cross-cutting modifiers:** Insurance-enhanced credit (Aon, Marsh, reinsurance) · NVIDIA + 6-sponsor platform (Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR). Tracked as quality tags on the sources above, not as additive categories. AI infrastructure 2026-28 · three headline measurements Third-party-associated financing stack. Excludes hyperscaler self-funded capex ($2.5-3.2T). $0.65T gross third-party financing of which $0.49T new money, $0.16T refi $0.95T formation + procurement $0.51T physical (L1-L3) + $0.44T compute (L4) $0.16T refi + secondary activity 24% of gross financing this window THESE ARE NOT ADDITIVE Each measures a different accounting quantity on a different basis. Summing them double-counts the same economic dollars. Modelled 5th-to-95th percentile ranges under common-factor Monte Carlo span wide bands above and below each figure. Bottom-up model · central estimates for 2026-28 window --- In 2026, AI infrastructure crossed a financing threshold. Risk that was previously concentrated on hyperscaler balance sheets is becoming contractually and financially modular without necessarily becoming economically independent. Claims on grid access, physical sites, equipment, compute cash flows, operator equity, and model-company demand are separated into different capital structures underwritten by different providers. Whether that separation constitutes genuine risk transfer, or only legal modularity that leaves economic exposure with hyperscaler balance sheets through take-or-pay, minimum-purchase commitments, and cross-recourse, is the research question the Risk Transfer Ratio (H.6) and Recourse-Adjusted Financing (H.7) metrics address quarterly in the Ledger. **Three third-party financing measurements for the 2026-28 window. Each measures a different quantity on a different accounting basis. None can be summed with the others.** - **Gross third-party financing: \~$0.65 trillion.** Of which $0.49T new money and $0.16T refinancing plus secondary flow. - **Third-party-associated formation + procurement: \~$0.95 trillion.** Split into two diagnostic components on different measurement regimes: $0.51T physical infrastructure formation L1-L3 (grid access, sites, electrical/cooling equipment) and $0.44T Compute & IT procurement L4 (revenue-derived proxy, not directly reconciled capex). - **Refinancing + secondary activity: \~$0.16 trillion.** Roughly 24 percent of gross financing this window. Identified external new-money capacity covers \~52 percent of the combined formation-and-procurement figure. The remaining 48 percent sits outside the Sources taxonomy: operator retained earnings, hyperscaler-adjacent structures, in-kind contributions, and other internal cash flows. Modelled 5th-to-95th percentile bands under common-factor Monte Carlo widen each figure materially. Two flows sit adjacent to this stack over the same window: - **Hyperscaler self-funded capex: \~$2.5-3.2 trillion.** A share overlaps the third-party stack through REIT leases and neocloud compute purchases. - **Model-company equity raises: \~$150 billion** primary plus secondary, of which perhaps $60-70 billion ultimately reaches infrastructure compute purchases downstream. The essay builds these measurements bottom-up from the physical constraints, with every input documented and every assumption graded on evidence quality. It also introduces the framework a recurring quarterly Ledger will use to track the capital stack as it evolves: not just how much capital flows, but where the risk sits, how the layers correlate, and where the systemic exposure builds. Each dollar of AI compute demand now creates a chain of financeable claims across five physical layers and five primary financing sources, connected by a small group of cross-layer capital providers, with a duration structure requiring continuous refinancing across technology generations. --- ## Model Boundary & Definitions **Methodological status.** This is a scenario model with accounting discipline, not a statistical forecast. Central estimates are model outputs based on documented inputs and stated assumptions, not observed market totals. Where inputs are Grade E (analyst scenario assumption), the model exposes them for replacement as evidence accumulates. The Monte Carlo output is a modelled distribution conditional on subjective input distributions and a common-factor correlation structure, not a probabilistic forecast. The current Monte Carlo approximates a networked dependency structure with a common-factor correlation; it does not yet model the stack as a full conditional-dependency graph. The Monte Carlo is therefore a model-uncertainty visualisation, not a probabilistic forecast of realised financing. Readers should evaluate the framework by whether its architecture holds and whether its assumptions can be independently refined, not by treating any central estimate as an empirical measurement. For rigour and to make the model reproducible, the following definitions apply throughout: - **MW basis.** All megawatt denominators in the layer calculations (the 65-142 GW range and every downstream figure) are expressed as IT-load-equivalent MW. Where source data are reported in interconnection or gross facility electrical MW, the model applies an upstream conversion (IT load MW = source MW × 0.65 to 0.85, central 0.75) to produce the IT-equivalent input range shown. The 0.65-0.85 factor represents PUE and source-definition differences (interconnection contracts and generation ratings do not equal IT load), not utilisation. The factor is Grade E and adjustable. It does not appear as a separate multiplier in any layer equation because the equations operate on already-normalised IT-load MW. - **Third-party-associated physical formation.** New physical infrastructure whose ownership, financing, procurement, or deployment is materially connected to a third-party financing structure. This is a scope definition; a given asset may be partially externally financed and still be included. - **New money.** Financing that funds newly-formed physical assets or new operator equity/debt facilities. Excludes refinancings of prior tranches, sale-leaseback recapitalisations of existing assets, and secondary equity sales. - **Refinancing.** Financing that recirculates against an already-financed asset or claim. - **Funding (for source taxonomy).** Includes debt, equity, customer prepayments, and vendor credit where the flow provides funding capacity to an infrastructure-owning or operating entity. Customer prepayments are treated as funding because they economically reduce the operator's external funding requirement, though they are not senior debt. - **External New-Money Coverage Ratio.** Identified external new-money capacity divided by third-party-associated formation and procurement. This is a capacity-to-uses ratio, not deal-level funding attribution: it measures the share identified as externally *coverable*, not "actually funded" (deal-level uses-to-sources allocation is not currently possible from public disclosure). "Coverage" rather than "intensity" or "penetration" because both of the latter imply deal-level allocation that the data does not support. - **Company equity valuations, including valuations of hyperscalers, AI labs and semiconductor vendors, are deliberately excluded from the financing Ledger.** Valuation is a stock measure of enterprise/equity value; the Ledger measures transaction flows and associated formation/procurement. The two should not be added or compared as equivalent capital quantities. §13 explains the reconciliation. - **Stress test vs falsifier.** A stress test evaluates whether the architecture holds under adverse conditions; a falsifier is evidence that would prove the architecture itself wrong. The essay distinguishes both. - **Cross-layer capital provider.** A capital participant holding material positions across three or more stack categories (physical layers, financing sources, upstream demand, or platform capital). Material threshold: more than $2 billion deployed or more than 10 percent of a single deal. Providers are counted at the **investment manager or affiliated capital platform** level. Blackstone counts as a single aggregation across its debt, growth-equity, real estate, and Tactical Opportunities funds. MGX counts as a single entity separate from the sovereign capital behind it. This prevents double-counting the same underlying pool through multiple co-investors, but also does not attempt to attribute exposure to ultimate beneficial owners (which is generally not disclosed). This is a research convention, not a market standard. --- ## 1\. Why the Stack Cannot Be Read as One Number Before doing the math, a methodological warning. The natural instinct is to add up the whole capital stack, produce a single trillion-dollar headline, and treat it as the size of the market. That instinct is wrong. It produces a number that is neither wrong nor right, but incoherent. Consider a single IREN-style transaction. IREN raises operator equity. That equity funds part of a GPU purchase. GPU-collateralised debt from Blue Owl funds most of the rest. The customer prepays roughly half the GPU capex against a compute offtake contract. The customer's ability to prepay comes from equity raised at the model-company layer. The model company's equity was raised in part from sovereign LPs who also participate in the neocloud sponsor's syndicate. Sum the flows and you can count the same dollar three or four times. Debt is not additive to the assets it finances. Sponsor equity is not additive to the debt structured against that sponsor's assets. Customer prepayments are not additive to the operator equity they complement. And the model-company equity that ultimately funds the customer's ability to prepay is a completely different unit of analysis from the GPU asset the prepayment supports. The right analytical response is to separate the accounting into three questions, each of which produces its own defensible number: **What is actually being built?** That is the physical uses side. Grid access, firm supply, sites, equipment, GPU capex. Measured in dollars of infrastructure formation, not dollars of financing activity. **Who finances it?** That is the capital sources side. Customer prepayments, vendor finance, GPU debt, real-estate and project finance debt, sponsor equity. Measured in dollars of financing flow, with an explicit split between new money and refinancing. **Where does the risk sit as the layers reprice?** That is the transmission side. A shock in one layer propagates to others with quantifiable propagation bands. Measured in cross-layer correlation, not in dollars. The rest of the essay works through each question with its own bottom-up construction. The three answers together describe the stack better than any single headline number could. Signature framework · introduced in the Ledger The Risk Transfer Test Two questions to distinguish. First: has the asset been legally separated from the hyperscaler balance sheet? That's usually visible in the filings. Second: how much economic downside has actually transferred with the asset? That's what the Risk Transfer Test measures quarterly. Nominally external financing → Recourse-adjusted financing (H.7) → True economic risk transfer (H.6) H.6 measures the share of formation and procurement for which material downside is contractually borne outside the hyperscaler/operator balance sheet. H.7 risk-weights nominal financing by the portion economically protected by hyperscaler take-or-pay, minimum-purchase, cross-recourse and strategic dependence. Baseline measurement in Ledger Vol 1 (December 2026). The Capital Stack in One Picture · uses ↔ sources ↔ modifiers Five physical layers on the left · five capital sources on the right · cross-cutting modifiers in the middle. PHYSICAL USES · $948BCAPITAL SOURCES · $646BL1a Grid access$73B central Interconnection deposits · transmission upgrades · generation triggered by AI-DC load L1b Firm supply$13B central SMR + advanced-reactor · BTM gas plant · AI-DC-attributable subset L2 Site + shell + civil$221B central Land · civil work · foundations · shell · security · fit-out L3 Equipment$204B central Transformers · MV switchgear · UPS · cold plates · PDUs · site-level backup L4 Compute & IT procurement$437B central GPU + accelerators + adjacent IT · non-hyperscaler subset (procurement proxy) S1 Customer prepayments$157B central Compute-offtake advances from model companies and hyperscalers S2 Vendor finance$30B central Equipment credit · extended payment terms retained by vendor S3 GPU-collateralised debt$224B central Senior debt secured against GPU fleets + offtake · Blue Owl-IREN print S4 RE / project-finance debt$140B central Site debt + BTM generation project finance · Digital Realty-Blackstone S5 Sponsor equity$95B central Growth equity + sovereign co-invest + listed operator issuance CROSS-CUTTING MODIFIERS(quality tags on sources, not additive)Insurance-enhanced creditAon · Marsh · Willis · reinsurancemakes previously unfinanceable risk financeableNVIDIA + 6-sponsor platform$500B announced · 20-50% conversioncontingent co-investor · sits inside S3/S4/S5Cross-layer capitalBlackstone · MGX · KKR · Brookfieldcorrelates layers through provider concentrationUses total $948B = $511B physical formation (L1-L3) + $437B compute procurement (L4). Sources total $646B = $490B new money + $156B refi/secondary. Not deal-level allocable. Anchor architecture for the AI Infrastructure Financing Ledger --- ## 2\. What Is Being Financed: The Physical Uses Side Five physical layers. Each measures dollars deployed to build a specific asset class, attributable to third-party financing structures, across 2026-28. **A note on layer count.** This essay uses five layers because that is the natural taxonomy for capital formation. The site's canonical AI Power Chain framework uses six engineering layers, running from grid at the top through substation, distribution, hall, and rack down to chip at the bottom. The financing model consolidates the chip-level layer into Compute (L4) alongside GPU capex, since financing structures for chip-level assets do not currently exist as a separate transactable category. The five-layer capital stack is not a revision to the six-layer engineering framework; it is a different taxonomy for a different purpose. ### Layer 1a. Grid access The senior physical layer. Interconnection deposits and development expenditure associated with securing grid access, transmission upgrades causally required by new AI-DC load, and generation additions triggered by the same load. Excluded: the utility's baseline reinforcement spend that would happen anyway, and queue positions themselves (which are administrative rights, not capital deployment). Bottom-up volumetric input across the six US regional transmission organisations plus international jurisdictions: - ERCOT: 25 to 45 gigawatts of AI-DC firm capacity reaching an executable interconnection pathway 2026-28\. The roughly 427-gigawatt AI-attributable queue would nominally clear 34 to 51 gigawatts at a historical queue-to-delivery benchmark of 8-12% (which is a general RTO observation from prior cycles, not an AI-specific forward conversion rate for the current queue composition); Governor Abbott's August audit directive delays some share of that. FERC's June 2026 show-cause orders (E-7 through E-12) directing large-load interconnection reform sit as the higher-level regulatory reset shaping how much of the queue can actually convert. - MISO: 10 to 20 gigawatts under the September 1 large-load framework's 120-day approval track (per MISO Large Load Working Group filings, April 2026). - PJM: 15 to 32 gigawatts across Northern Virginia, Ohio and the mid-Atlantic corridor. - Other US ISOs (CAISO, ISO-NE, NYISO, SPP): 5 to 15 gigawatts combined. - International: 10 to 30 gigawatts including Norwegian hydro, French and German markets, Nordic, UK, Japan, Singapore, Australia, Korea. Global range: 65 to 142 gigawatts of AI-DC firm capacity clearing during the window, central estimate 104 gigawatts. Capital deployment at Layer 1a is the sum of interconnection deposits plus development capital plus utility and IPP upgrade capex specifically attributable to serving new AI-DC load. The last item is the largest and the most methodologically sensitive. The attribution question (what share of a utility's incremental capex is caused by AI-DC load rather than baseline system reinforcement) has no clean empirical answer. The model uses a central attribution rate of 60 percent as the working assumption because a substantial share of incremental infrastructure in large-load projects is directly required to serve the new load; the empirical attribution rate remains uncertain. The full range spans 40 to 80 percent. The capital intensity inputs are $60-140/kW for interconnection deposits and development expenditure (central $100/kW), and $600-1,400/kW for gross transmission and generation upgrades (central $1,000/kW), of which 40-80 percent (central 60 percent) is attributable to new AI-DC load rather than baseline reinforcement. Bottom-up: - Low: 65 GW × ($60/kW + $600/kW × 40%) = 65 × $300/kW = **$19.5 billion** - Central: 104 GW × ($100/kW + $1,000/kW × 60%) = 104 × $700/kW = **$72.8 billion** - High: 142 GW × ($140/kW + $1,400/kW × 80%) = 142 × $1,260/kW = **$178.9 billion** Layer 1a contributes roughly **$20 to $180 billion**, central approximately **$73 billion**. ### Layer 1b. Firm supply Distinct from grid access because it measures generation capex that delivers electrons rather than the transmission that carries them. Two sub-components remain physical formation; a third (PPA prepayments) is tracked separately as a contractual financing mechanism. **Physical formation (in L1b total):** - Small modular reactor and advanced-reactor development capital attributable to AI-DC offtake. X-energy received a further $1 billion of DOE cost-share funding for Long Mott on August 17\. TerraPower signed Hyundai Engineering & Construction as EPC for eight Natrium reactors on the same day. AI-DC attribution rule: project capex × AI-DC contracted capacity share of the reactor's offtake portfolio (currently 20-40 percent at the projects tracked). Range: $2-9 billion, central $5 billion. - Behind-the-meter gas plant financing. Global Energy Monitor's August tally of US gas-fired capacity in the AI-DC pipeline reached 189 gigawatts across announced, pre-construction, and construction stages. This is a project pipeline, not committed generation capacity. The subset that is third-party-associated and commissioned inside 2026-28 is estimated at 5-18 gigawatts at $500-1,100/kW. Range: $2.5-19.8 billion, central $7.5 billion. **Firm supply (L1b) physical formation total: $5-29 billion, central $13 billion.** **Contractual (not in Firm supply (L1b), tracked separately):** - PPA prepayments and capacity reservation fees paid by hyperscalers and neoclouds to secure firm capacity. Template: the Google-Kairos-TVA agreement announced August 18 for 50 megawatts of Kairos Hermes 2 SMR capacity. These are financing/contractual arrangements that support physical capex rather than being physical capex themselves. Estimated at $2-11 billion over the window, central $5 billion. They appear as a financing signal in the Ledger but do not enter the physical-uses total. **PPA prepayments are not included in Source 1 unless specifically applied to GPU or compute procurement; they remain a separate contractual-capital indicator to avoid double-counting future Ledger flows.** ### Layer 2\. Site, shell and civil works This is where v1 of this model conflated site construction with electrical and cooling equipment. The rebuilt taxonomy separates them cleanly. Layer 2 includes only: - Land acquisition - Civil site work, grading, foundations - Building shell, roofing, structural steel - Security perimeter, access control infrastructure - Office fit-out and operations centre Explicitly not included in Layer 2, reassigned to Layer 3: main electrical transformer, medium voltage switchgear, low voltage panels, uninterruptible power supplies, coolant distribution units, cold plates, power distribution units, backup generators. The narrower Site+shell (L2) definition produces a narrower per-megawatt cost range: $1.8 to $3.6 million per megawatt for site plus shell plus civil, central $2.6 million. Compared to industry benchmarks of $10 to $15 million per megawatt for a fully-fitted-out AI data centre including all electrical and cooling equipment plus IT capex, this places Site+shell (L2) at roughly 20 to 30 percent of the total build cost, which matches published REIT and private-sponsor project economics. Bottom-up. The 60-90 percent factor is a **site-formation conversion**: the share of AI-DC firm capacity that reaches site and civil formation (foundation poured, shell rising or complete) within the 2026-28 window, on IT-load-equivalent MW terms. **This measures physical site formation, meaning the point at which foundations are poured and shell is rising or complete. Site formation is a substantially earlier milestone than delivery and commissioning.** - Site build low: 65 GW × 60% site-formation conversion × $1.8M/MW = **$70.2 billion** - Site build central: 104 GW × 75% site-formation conversion × $2.6M/MW = **$202.8 billion** - Site build high: 142 GW × 90% site-formation conversion × $3.6M/MW = **$460.1 billion** Plus sale-leaseback new-formation portion of $10-30 billion, central $18 billion (see below). **Site+shell (L2) physical formation total: $80-490 billion, central $221 billion.** Adjacent to but not identical to Site+shell (L2) formation: sale-leaseback secondary activity in the data-centre real-estate market. The Digital Realty to Blackstone $7.8 billion Northern Virginia transaction, announced June 2026 per Digital Realty press release, is the observed pricing/comparable print for the sub-market. The $10-30 billion new-formation figure below is the modelled aggregate 2026-28 new-formation component of sale-leaseback activity across the market, not the amount in that single transaction. Most of this transaction represents recapitalisation of existing assets rather than new formation. The portion that funds new build is estimated at $10 to $30 billion during 2026-28, central $18 billion. The rest, new-formation excluded, flows into Number 3 (refinancing and secondary activity) rather than Number 2 (formation and procurement). ### Layer 3\. Electrical and cooling equipment **Methodological note.** Equipment (L3) is a vendor-flow triangulation from disclosed backlog, new orders, and above-backlog demand, not a project-level capex census. It sizes the aggregate equipment expenditure supporting AI-DC formation over the window rather than reconciling to any single project. Everything in the physical build that is not shell and not IT. The full taxonomy: - Main transformer stepping utility feed to medium voltage - Medium voltage switchgear and protection - Low voltage transformers and panels - Uninterruptible power supplies - Coolant distribution units - Cold plates and direct-liquid-cooling manifolds - Power distribution units at rack level - Backup generators (increasingly reciprocating gas units) Sizing draws from vendor Q2 2026 backlogs, with an important boundary correction from v7\. GE Vernova's utility-scale turbine backlog (116 GW to 2031, of which $60-80 billion is AI-DC-attributable) is generation equipment and properly belongs in Layer 1b (firm supply). Only the site-level backup-generation share of GE Vernova and other genset vendors ($10-20 billion) is included in Equipment (L3) to avoid Firm-supply/Equipment double-counting. Vendor line items in Equipment (L3) (all AI-DC-attributable, third-party-associated subset): - Siemens Energy Grid Technologies: €51 billion (\~$55B) reported order backlog per H2 FY25 filing; estimated AI-DC attributable share 60-75% = $33-41 billion (author's model) - Vertiv: reported >$15 billion, estimated $15-18 billion - Eaton (post-Boyd Thermal integration): $12-15 billion - Schneider Electric: $10-12 billion - ABB: $8-10 billion - Hitachi Energy: $8-10 billion - Delta Electronics: $5-7 billion - Legrand + Mitsubishi Electric + TDK + other rack/modular: $5-8 billion - GE Vernova + Cummins + Wärtsilä site-level backup generation only: $10-20 billion - Cooling specialists (Ecolab-CoolIT, Trane-LiquidStack, Nautilus, Motivair, Rittal, Asetek): $8-12 billion - Prefab modular power infrastructure (Vertiv MegaMod, Schneider PowerModule, Silverdraft): $8-15 billion - Chinese and other inverter/UPS/switchgear vendors serving AI-DC (Sungrow, Envision, Kstar): $15-25 billion - Systems integration and balance-of-plant electrical: $10-25 billion Sum of the vendor line items above: $137 to $208 billion, central $173 billion. **These figures are the third-party-associated subset; the 55-75 percent third-party filter is applied at the vendor-line-item level (each vendor's AI-DC-attributable backlog × third-party share), not layered as an additional multiplier on the sum.** Bottom-up L3 formation: | L3 component | Low ($B) | Central ($B) | High ($B) | Method | | ---------------------------------------------------- | -------- | ------------ | --------- | ------------------------------------------------------------------------------------------------------------------ | | Existing backlog conversion 2026-28 (Grade D) | 69 | 112 | 162 | Backlog $137-208B × conversion 50-78% | | In-window new orders, fast-ship categories (Grade E) | 33 | 72 | 128 | Cooling + PDU + some UPS placed and shipped in 2026-27, with same 55-75% third-party-association factor as backlog | | Above-backlog cooling uplift (Grade E) | 8 | 20 | 38 | Reasoning-model thermal density + storage cooling extension, third-party-association filtered | | **L3 total** | **110** | **204** | **328** | Sum | Layer 3 · vendor-flow triangulation to $204B central Not a project-level capex census. Sizes aggregate equipment expenditure supporting AI-DC formation. $0B$50B$100B$150B$200B+$112BExisting backlogconversion+$72BIn-windownew orders+$20BAbove-backlogcooling uplift$204BL3 centralTotal L3 Backlog range $137-208B (central $173B) × 55-75% third-party filter × 50-78% conversion + new orders + cooling uplift Equipment (L3) formation across 2026-28 combines backlog conversion (50 to 78 percent of August backlog shipping in-window, given Siemens Energy's own guidance that capacity expansion will not fully arrive until 2030), new orders placed and shipped in the same window (fast-ship categories including cooling and PDUs), and above-backlog demand from reasoning-model thermal density. ### Layer 4\. Compute & IT procurement **Methodological note.** Compute (L4) should be read as a scenario-weighted procurement proxy derived from vendor revenue rather than a bottom-up estimate of third-party GPU capex. Two Grade-E assumptions (non-hyperscaler share and hyperscaler-adjacent overlap) jointly determine a large share of the central. Third-party financed GPU and adjacent IT procurement, excluding hyperscaler self-funded purchases. Largest layer by dollar contribution. Labelled "procurement proxy" rather than "physical formation" because the underlying inputs (NVIDIA data-centre revenue and comparable) are revenue/procurement figures, not directly reconciled capex bridges; the label prevents the conceptual mismatch between revenue-side proxy and physical-formation denominator. Bottom-up build: | L4 component | Low ($B) | Central ($B) | High ($B) | Method | | ------------------------------------------------------- | -------- | ------------ | --------- | ----------------------------------------------------------------------------- | | NVIDIA data-centre revenue to non-hyperscaler operators | 325 | 495 | 680 | Total NVIDIA DC revenue 2026-28 est. $1.3-1.7T × 25-40% non-hyperscaler share | | AMD + other accelerator capex through third parties | 40 | 70 | 120 | MI-series + Blackwell-adjacent through neocloud operators | | Adjacent IT (networking, storage, integration) | 30 | 50 | 80 | \~8-10% of GPU + accelerator capex for networking + storage tiers | | Less: hyperscaler-adjacent scope overlap | \-79 | \-178 | \-308 | Overlap 20% low / 29% central / 35% high applied to pre-overlap subtotal | | **L4 total** | **316** | **437** | **572** | Sum | Layer 4 · scenario-weighted procurement proxy build-up Revenue-derived. Not a bottom-up estimate of third-party GPU capex. $0B$200B$400B$600B$700B+$495BNVIDIA DC rev ×non-hyperscaler+$70BAMD + otheraccelerators+$50BAdjacent IT(net + storage)\-$178BHyperscaler-adjacentscope overlap$437BL4 centralCompute & ITprocurement NVIDIA DC revenue proxy × non-hyperscaler share (33%) less hyperscaler-adjacent overlap (29%) **Derivation of the scope-overlap adjustment.** NVIDIA data-centre revenue used as a revenue-side proxy for accelerator and associated system procurement (not as a direct measure of GPU capex) inevitably includes chips that ultimately sit inside hyperscaler-funded infrastructure through joint-venture, colocation-anchored, or neocloud-serving-hyperscaler-workload arrangements. Scope overlap is applied at 20 percent (low) / 29 percent (central) / 35 percent (high) of each column's pre-overlap subtotal: low case $395B × 20% = $79B; central $615B × 29% = $178B; high $880B × 35% = $308B. Calibrated to reported neocloud capacity contracted to hyperscaler workloads and related scope-overlap evidence rather than directly observed at portfolio level. The adjustment removes activity that is economically within the hyperscaler-direct scope, rather than correcting a literal accounting duplication within NVIDIA's revenue. Two dominant uncertainties: how much of NVIDIA's data-centre revenue is ultimately deployed by neoclouds and independent operators versus by hyperscalers directly, and how the Rubin generation shipment schedule during 2027-28 tracks against announced roadmaps. **NVIDIA revenue caveat.** NVIDIA data-centre revenue is used as a revenue-side proxy for accelerator and associated system procurement, not as a direct measure of GPU capex. The Adjacent IT line above measures non-NVIDIA networking, storage, and integration to avoid internal double-counting. ### Physical uses total **Evidence grades:** A sourced transactionB company disclosureC industry datasetD analyst inferenceE scenario assumption Bottom-up reconciliation of physical uses (central values, rounded from model outputs): | Layer | Low ($B) | Central ($B) | High ($B) | Grade | | ----------------------------------------------------------- | -------- | ------------ | --------- | ----- | | L1a Grid access | 20 | 73 | 180 | DE | | L1b Firm supply | 5 | 13 | 29 | DE | | L2 Site + shell + civil | 80 | 221 | 490 | DE | | L3 Equipment | 110 | 204 | 328 | D | | **Physical infrastructure formation (L1-L3)** | **215** | **511** | **1,027** | mixed | | L4 Compute & IT procurement | 316 | 437 | 572 | CE | | **Combined third-party-associated formation + procurement** | **531** | **948** | **1,599** | mixed | Physical uses side · $B deployed 2026-28 Range = low-to-high scenario. Marker = central estimate. $0B$100B$200B$300B$400B$500B$600BL1a Grid access$73B$20-180BL1b Firm supply$13B$5-29BL2 Site + shell + civil$221B$80-490BL3 Equipment$204B$110-328BL4 Compute & IT procure$437B$316-572BPhysical infrastructure formation (L1-L3): $215-1,027B, central $511BCombined formation + procurement: $531-1,599B, central $948B Five-layer bottom-up build · L1a grid access · L1b firm supply · L2 site+shell · L3 equipment · L4 compute Central rounds to approximately **$0.95 trillion** of third-party-associated formation and procurement across 2026-28: $511 billion of physical infrastructure formation (L1-L3) and $437 billion of compute and IT procurement proxy (L4). Layer values may not sum exactly to the total in unrounded terms because model calculations use non-integer intermediate values; the reported total uses summed layer centrals rather than any independent aggregate. The wide range is honest. It reflects that four of the five layers have inputs graded D or E on evidence quality, meaning estimated with method or scenario-only. Tightening the range requires better quarterly disclosure from utilities on AI-attributable capex, from operators on unit-economics of buildable sites, and from equipment vendors on AI-DC-specific backlog conversion. Each of those improvements will land over the next 12 months and become inputs into the recurring Ledger. --- ## 3\. Who Funds It: The Capital Sources Side Five primary sources fund the physical uses above. Two additional categories (insurance-enhanced credit and NVIDIA-platform capital) are tracked as cross-cutting quality modifiers rather than additive sources, since they attach to deals originated under the primary sources rather than originating capital independently. ### Source 1\. Customer prepayments IREN's FY2026 Q4 earnings disclosure (released 27 August 2026, ahead of the Blue Owl financing announcement on 28 August 2026) states that recent customer prepayments represent 45 to 55 percent of associated GPU capex on the Blue Owl-financed Blackwell Ultra tranche. On that math, customer prepayments are the largest single source of capital funding third-party GPU purchases in the current cycle. Applied to the Layer 4 accelerator/GPU procurement subset (approximately 75-85 percent of the L4 procurement proxy total; the balance is networking/storage/integration which is typically not prepayment-supported), with a blended prepayment share of 30 to 55 percent (lower than IREN's 45-55 percent because non-investment-grade-anchored deals typically get less prepayment support), customer prepayments contribute: - Low: $316B L4 × 75% GPU subset × 30% prepayment = **$71B** - Central: $437B L4 × 80% GPU subset × 45% prepayment = **$157B** - High: $572B L4 × 85% GPU subset × 55% prepayment = **$267B** Range: **$71-267B, central $157B**. The prepayment ratio applies to the accelerator subset rather than to the full L4 procurement proxy: non-accelerator L4 categories are not typically prepayment-supported at the same rate. The mechanism matters. A customer prepayment is capital that flows from the model company or hyperscaler counterparty directly to the operator, bypassing the debt and equity markets. It reduces the operator's external financing requirement. It reduces the operator's net external funding requirement and can improve lender protection by reducing the amount of capital that must be funded against the same contracted deployment. Economically, the prepayment transfers part of the asset-funding burden to the customer before debt is raised. Prepayment support is one factor that can improve lender protection and help explain why a structure of this type can clear at a tighter price than an otherwise comparable fully externally funded neocloud, though the counterfactual pricing is not observable. Customer prepayments are likely an important enabling condition for the scale and pricing of the current third-party GPU financing market. Without them, the market would plausibly be smaller or priced considerably wider, though the counterfactual is not directly observable. ### Source 2\. Vendor finance Equipment credit and extended payment terms where economic exposure is retained by the vendor or explicitly structured as vendor credit. Distinct from the NVIDIA plus six-sponsor platform (co-invested capital, classified under Source 3-adjacent structures) and from bank refinancing of vendor receivables (classified under the relevant debt source). Range: $15 to $55 billion, central $30 billion. Small in absolute terms but a leading indicator of vendor confidence in cycle durability. ### Source 3\. GPU-collateralised debt Senior debt secured against GPU fleets plus underlying offtake contracts. Sized from advance rate times addressable capex, split between new money and refinancing. The Blue Owl to IREN deal announced August 28 is the pricing print. Structure: $1.2 billion senior secured term loan plus $1.2 billion senior secured notes, 9 percent coupon, 2.5-year tenor, PIMCO participating. IREN disclosed that this financing represented approximately 90 percent of associated GPU capex on the funded tranche, with the balance covered by customer prepayments and operator equity. That implies an advance rate at the higher end of the range for well-structured, prepayment-supported deals. The essay uses a portfolio-level advance rate range of 55 to 82 percent, reflecting the mix of high-quality, prepayment-supported deals like IREN's and lower-quality neocloud debt. Applied to a bankable-capex base of $200 to $500 billion, gross GPU debt origination during the window is $110 to $410 billion, central $224 billion. Split between new money and refinancing: the first-generation cycle skews heavily toward new money. Central new-money share is 72 percent, giving new money of $161 billion and refinancing of $63 billion. As the market matures and first-tranche 2.5-year debt approaches maturity in late 2028, the refinancing share rises. Two important points on the 9 percent coupon. First, it prices a bundle. The roughly 400 basis point spread over a comparable-tenor investment-grade hyperscaler benchmark (typical 2.5-year to 3-year investment-grade hyperscaler debt currently in the \~5-5.5 percent range) does not isolate GPU credit risk. The 9 percent coupon prices operator credit, customer concentration, contract quality, GPU residual value, structural subordination, collateral enforceability, and private-credit illiquidity all at once. The correct read is that this is the first observable market price for a financing structure in which GPU collateral, compute offtake and operator credit jointly underwrite the debt, not that GPUs specifically carry a 400 basis point risk premium. Second, one transaction does not establish an asset-class discount rate. It establishes a first pricing print. The asset class becomes genuinely priced when there are multiple transactions with different counterparties, different tenors, different collateral, different structures, and observable dispersion. The Ledger will track that dispersion as subsequent deals close. ### Source 4\. Real estate and project finance debt Debt secured against data-centre sites and against BTM generation projects. Distinct from Source 3 because collateral is real estate or long-lived generation asset rather than GPU fleet. Includes both direct debt origination against DC sites and secondary transactions like the Digital Realty-Blackstone Northern Virginia deal, reported gross transaction value $7.8 billion including assumed debt and remaining capex per Digital Realty press release; Digital Realty consideration for Blackstone's interests approximately $3.5 billion. Range: $80 to $220 billion, central $140 billion. Split between new formation (40 to 70 percent) and refinancing / sale-leaseback recapitalisation (the balance). At the central mid-point, new money is $77 billion and refinancing is $63 billion. ### Source 5\. Sponsor equity (primary/secondary split) Growth equity, sovereign co-investment sleeves, and listed operator issuance funding neocloud and independent operator platforms. New money excludes secondary equity sales per the Model Boundary definitions; the model now splits sponsor equity (S5) gross into primary (new money) and secondary components: - **S5 gross: $55-145B, central $95B** (includes primary and secondary) - **S5a Primary sponsor/operator equity (new money): $40-100B, central $65B** - **S5b Secondary equity (moves to Number 3 refi/secondary): $15-45B, central $30B** Only S5a enters the new-money aggregate. S5b is captured in the refinancing/secondary total. Includes: - Firmus reportedly raised $2 billion at $10.5 billion post-money (August 7, per press coverage), Coatue-led with NVIDIA, Blackstone Tactical Opportunities and Jane Street - CoreWeave secondary and follow-on issuance - IREN, Nebius listed equity flow - Growth equity dry powder deployment from Coatue, Blackstone Growth, Sequoia, Andreessen Horowitz Machine Age Fund ($1.1 billion closed August 28), plus sovereign co-invest sleeves - APAC and EU operator equity beyond the primary US flow Range: $55 to $145 billion gross, central $95 billion. Primary/secondary split shown below. Sovereign LP share of sponsor equity is estimated at 25 to 55 percent, central 40 percent. This is a tracking metric for the recurring Ledger (sovereign concentration in operator equity is one of the strongest signals of cycle durability). ### Source 6\. Insurance-enhanced credit Not a separate financing source. A quality tag on the debt sources above. Aon Data Center Lifecycle Insurance Program at $5 billion, up from $3.5 billion in April and $2.5 billion in January, is the disclosed anchor point. Additional programmatic capacity from Marsh, Willis Towers Watson, Munich Re, Swiss Re, and Lloyd's syndicates exists but is not fully disclosed at the programmatic level; the Ledger will track disclosed capacity from these markets as it becomes public. Total AI-DC-specific insurance capacity in the current market is best characterised as a growing but partly opaque quantity, not a single quantifiable pool. The insurance capacity is important not because it directly funds infrastructure but because it makes previously unfinanceable risk financeable. Construction all-risks coverage, delay-in-start-up coverage and business interruption coverage together transform the credit profile of the debt sources above from concentrated tech-cycle risk to project-finance-adjacent risk. Aggregate disclosed insurance capacity is a leading indicator of whether the debt sources above can continue to expand at current pricing. ### Source 7 (as tracking metric only). NVIDIA plus six-sponsor platform Announced August 10 at $500 billion of capacity across NVIDIA plus Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. NVIDIA is disclosed to hold a minority contingent co-investor role in transactions structured under the platform. NVIDIA is not the borrower; sponsors are the principal capital. Announced capacity is not deployed capital. The model assumes 20 to 50 percent of announced capacity translates into closed third-party transactions during 2026-28, central 30 percent; this is a scenario assumption rather than an observed conversion rate. Applied to $500 billion of announced capacity, that produces $100 to $250 billion of platform-structured deals during the window, central $150 billion. Held as a tracking metric rather than an additive source because platform deals sit inside Sources 3, 4, or 5 depending on deal structure. The Ledger will track platform deployment share as a proxy for cycle maturity, and will track whether the platform becomes a dominant financing mechanism as it scales. ### Capital sources total Sources 1 through 5, summed without double-count adjustment: - Gross third-party financing activity 2026-28: $328 to $1,097 billion, central $646 billion - New-money portion: $220 to $959 billion, central $490 billion - Refinancing/secondary portion (includes S5 secondary equity): $27 to $583 billion, central $156 billion Capital sources · gross financing 2026-28 Five primary sources. Insurance + NVIDIA platform tracked as cross-cutting modifiers, not additive. $0B$50B$100B$150B$200B$250B$157BS1 Customer prepayments$30BS2 Vendor finance$224BS3 GPU-collat debtnew $161 · refi $63$140BS4 RE/PF debtnew $77 · refi $63$95BS5 Sponsor equitynew $65 · refi $30New money ($490B central, 76% of gross)Refi + S5 secondary ($156B, 24% of gross)Gross third-party financing: $328-1,097B, central $646B \= $490B new + $156B refi/secondary Bottom-up model · new-money vs refi split per Model Boundary The gap between the deterministic range and the common-factor Monte Carlo output matters. A 10,000-run common-factor Monte Carlo simulation samples each input as Xᵢ = 0.6·F + 0.4·εᵢ where F is a common demand-shock factor sampled from a truncated normal and εᵢ is an independent draw from each input's triangular distribution. The 0.6 is the common-factor loading (not a pairwise correlation target). This produces a much tighter modelled distribution than the naive low-to-high range implies. The common-factor structure captures broad co-movement but does not represent the full conditional dependence structure between financing variables; some pairs of variables may correlate negatively rather than positively (for example, financing spreads and lender advance rates may move inversely to demand). Gross financing 5th to 95th percentile under the model's correlated uncertainty assumptions is $446 to $1,021 billion, mean $690 billion. New-money financing 5th to 95th percentile under the same assumptions is $333 to $883 billion, mean $560 billion. The mean differs slightly from the deterministic central because the simulation samples the full input triangular distributions rather than fixing each input at its displayed central value. Two intellectual disclaimers about the Monte Carlo output that a sophisticated reader will need. First, the modelled percentile range should not be read as a statistical confidence interval. It is a conditional simulation range, dependent on the subjective input distributions and the common-factor correlation assumption. A statistical confidence interval requires observed data with known distributional properties. This model has neither. The percentile range describes what the model produces, not what reality is. Second, the effective dependence structure is likely stronger than the 0.6 common-factor representation captures in some states of the cycle. The layers are economically linked at every stage: an AI demand shock hits model-company equity first, then customer prepayments, then GPU debt, then sponsor equity, then operator commitments, then equipment orders, then site construction, then grid development. Some pairs of variables (financing spreads vs. lender advance rates, for instance) may move inversely to each other in ways that a single positive common factor cannot represent. Independent-sampling Monte Carlo systematically understates tail risk in the presence of positive co-movement; a single positive common factor partially fixes that but cannot represent conditional or sign-changing dependence. The right reading of the Monte Carlo output: the central estimate is the model's best current estimate rather than an observed market quantity; the distribution around it is a rough guide to model uncertainty conditional on the input distributions; the actual outcome distribution is likely wider than any of these numbers imply, particularly in the tails. **The practical takeaway.** For an underwriter or capital allocator: treat the central estimates as best-current-guesses. Treat the modelled percentile bands as a floor on real-world uncertainty; actual outcomes are likely wider in the tails. The model earns its keep through the Ledger's quarterly refresh of the highest-sensitivity inputs (L4 non-hyperscaler share and hyperscaler-adjacent overlap; L2 site-formation conversion; IT-load conversion factor). Track the direction and magnitude those inputs move quarter to quarter, and whether the transmission bands hold when actual shocks land. That is the observable signal that matters for underwriting. Gross financing · common-factor Monte Carlo distribution Modelled distribution conditional on subjective input distributions and 0.6 common-factor loading. $400B$600B$800B$1000B$1200BP5 · $446BMean · $690BP95 · $1,021BDeterministic central $646B Not a probabilistic forecast. Approximates networked dependency with common-factor correlation; not a full conditional-dependency graph. --- ## 4\. External New-Money Coverage Ratio: Reading the Gap Between Uses and Sources Attentive readers will notice that the combined third-party-associated formation plus procurement total (central $948 billion) is larger than the sources total (central $646 billion). The gap is a scope difference, not an accounting error. Reading it correctly gives one of the most useful ratios in the entire framework. Uses measure third-party-associated formation and procurement attributable to the financing stack. Sources measure explicitly identified external-financing instruments. But a third-party-owned asset can be funded partly by external financing and partly by capital that does not appear in the Sources taxonomy: - Operator retained earnings from existing compute contracts, reinvested into capacity expansion - Hyperscaler capex that flows through joint-venture or partnership structures where the physical asset is third-party structured but the funding source is hyperscaler balance sheet - Sale-leaseback proceeds that fund new formation while being categorised in the recapitalisation bucket for accounting purposes - Customer purchases of compute at contracted prices that generate operator cash flow, some of which cycles back into capacity investment The gap therefore does not indicate a financing shortfall. It reflects that the Source taxonomy captures external financing instruments, while internal cash flow, hyperscaler-adjacent structures, and vendor-in-kind financing sit outside that taxonomy. The ratio to track is identified external new-money capacity relative to third-party-associated formation and procurement, which we call the **External New-Money Coverage Ratio**, rather than the absolute gap itself. This is a capacity-to-uses ratio, not deal-level funding attribution. At central estimates, using new-money financing (not gross financing, which includes refinancing that does not fund new formation) against third-party-associated formation and procurement: | Metric | Numerator | Value | % of combined formation + procurement | | --------------------------------- | ------------------------------ | ----- | ------------------------------------- | | External New-Money Coverage Ratio | Total new-money sources | $490B | **52%** | | New-money debt intensity | S3 new + S4 new ($161B + $77B) | $238B | **25%** | | Customer-funded intensity | S1 customer prepayments | $157B | **17%** | | Sponsor-equity intensity | S5a primary sponsor equity | $65B | **7%** | | Vendor-finance intensity | S2 vendor finance | $30B | **3%** | External New-Money Coverage Ratio · $490B / $948B Identified external new-money capacity as share of third-party-associated formation + procurement. 52%Coverage RatioNew-money debt (S3+S4)$238B · 25% of $948BCustomer prepayments (S1)$157B · 17% of $948BPrimary sponsor equity (S5a)$65B · 7% of $948BVendor finance (S2)$30B · 3% of $948BUnidentified in taxonomy$458B · 48% of $948B Not a deal-level funding attribution. Capacity-to-uses ratio; deal-level allocation unavailable. The individual categories sum to 52 percent, matching the aggregate Coverage Ratio. The interpretation: identified external new-money capacity equals approximately 52 percent of third-party-associated formation and procurement at model central. The remaining 48 percent is not identified in the current Source taxonomy. It could be funded by operator retained earnings, hyperscaler-adjacent structures, in-kind contributions, other internal cash flows, or external sources not separately identified in the model; the model does not assert the composition, only that it is not captured in the identified Sources. A separate ratio: **refinancing share of gross financing = $156B / $646B = 24 percent.** Roughly one dollar in five of third-party financing activity in the current window recirculates rather than funds new formation. That share should rise materially as the first-generation refinancing window opens in late 2028, although the timing and concentration of maturities remain to be demonstrated empirically in the Ledger. These are the ratios the Ledger will track each quarter. Their evolution matters more than the absolute totals. Rising debt intensity signals financialisation. Rising customer-funded intensity signals demand-side concentration. Rising sponsor-equity intensity signals cycle immaturity or heightened risk aversion. Rising refinancing share signals cycle maturity. Each trajectory carries a different implication for cycle durability. Sources → Uses · flows are illustrative, not deal-level attribution Aggregate capacity-to-uses view. Curves do not represent actual deal-level funding flows. L1a Grid$73BL1b Firm$13BL2 Site+shell$221BL3 Equipment$204BL4 Compute$437BS1 prepay$157BS3 GPU debt (new)$161BS4 RE/PF debt (new)$77BS5a primary eq$65BS2 vendor$30B$458B unidentifiedoperator retained earnings + hyperscaler-adjacentstructures + in-kind + other internal cash flowsSOURCES · $490BUSES · $948B Public disclosure does not currently support deal-level uses-to-sources allocation --- ### The capital recycling mechanism IREN's FY2026 Q4 disclosure surfaces something more interesting than any single financing ratio. The company reports that recent customer prepayments plus GPU debt combined can exceed 100 percent of associated GPU capex on the narrow "GPU purchase" denominator. The interpretation matters: this does not mean the same asset has been financed more than once. It means the financing package provides funding capacity greater than the narrow GPU purchase-price denominator because the financing supports a broader eligible-use and collateral pool (adjacent site, electrical infrastructure, and working capital tied to the same deployment). The mechanism is capital extension across uses, not double-counting. A well-structured financing package sized to a narrow GPU tranche can support the GPU plus adjacent site and electrical infrastructure required to deploy the GPU, with the exact allocation depending on the operator's use of proceeds. That is a legitimate feature of structured infrastructure finance. The framework implication: financing sources can cross-fund uses. Customer prepayments raised against a specific compute contract can partly fund the GPU that services that contract and partly fund adjacent physical formation. This is why the Sources total is smaller than the Uses total in absolute terms but represents a larger effective financing base once cross-funding is counted. The Ledger will track this metric with a further distinction: **committed funding coverage** (committed debt + committed equity + contractual prepayment divided by eligible capex) vs **funded coverage** (drawn debt + funded equity + received prepayment divided by incurred capex). The distinction matters because undrawn commitments and unreceived prepayments are not deployed capital until they fund. Ratios above 100 percent, as IREN currently demonstrates on the narrow GPU denominator, signal that financing packages are generating capacity beyond the immediate asset purchase. Ratios well below 100 percent signal financing stress. **Funding coverage** is distinct from the **External New-Money Coverage Ratio**, which the Ledger separately tracks as identified new-money financing divided by third-party-associated formation and procurement. Both matter and measure different things. Funding coverage is a deal-level capacity ratio (individual-deal financing headroom against its own eligible capex). The External New-Money Coverage Ratio is an aggregate capacity ratio (identified external new-money capacity against aggregate modelled uses). Neither is a deal-level uses-to-sources allocation, which public disclosure does not support. The term "capital velocity" is reserved for cumulative financing flows divided by average deployed capital (a proper turnover measure), which the Ledger may track separately as a future metric once the outstanding-capital denominator is properly calibrated. --- ## 4a. Model Audit Table The dominant inputs to the $0.95T central estimate, with evidence grade, formula, double-count risk, and sensitivity to reasonable variation. This is the target-of-attack surface for anyone stress-testing the model. | Input | Central | Low | High | Grade | Formula | Double-count risk | Sensitivity | | ----------------------------------- | ------- | ---- | ------ | ----- | ----------------------------------------------------------------- | --------------------------------------------------- | ----------- | | Firm capacity (IT-load MW) | 104 GW | 65 | 142 | D | ERCOT 25-45 + MISO 10-20 + PJM 15-32 + other US 5-15 + intl 10-30 | Low (bounded by IT-MW normalisation) | Very high | | IT-load conversion factor | 0.75 | 0.65 | 0.85 | E | IT MW = source MW × factor | High (PUE/definition drift) | Very high | | L1a AI attribution | 60% | 40% | 80% | E | of gross utility/IPP upgrade capex per new GW | Medium (baseline reinforcement) | High | | L1a $/kW gross upgrade | $1,000 | $600 | $1,400 | D | utility rate case + IPP disclosures | Low | High | | L2 site-formation conversion | 75% | 60% | 90% | E | share of firm capacity reaching site/civil formation in window | Medium (with L1a) | Very high | | L2 $/MW site+shell+civil | $2.6M | $1.8 | $3.6 | D | REIT capex disclosures normalised to IT MW | Low (excludes L3) | Very high | | L3 backlog (third-party-associated) | $173B | 137 | 208 | D | vendor Q2 2026 disclosures × 55-75% third-party share | Medium (GE Vernova boundary with Firm supply (L1b)) | High | | L3 backlog conversion rate | 65% | 50% | 78% | D | typical vendor backlog-to-shipment cycle | Low | High | | L4 NVIDIA DC revenue 2026-28 | $1.5T | 1.3 | 1.7 | C | current run rate × growth guidance | Low | Very high | | L4 non-hyperscaler share | 33% | 25% | 40% | E | Nvidia disclosure + operator triangulation | High (definitional) | Very high | | L4 hyperscaler-adjacent overlap | 29% | 20% | 35% | E | neocloud-serving-hyperscaler share, calibrated not observed | Very high | Very high | | S1 customer prepay share of GPU | 45% | 30% | 55% | A/B | IREN FY26 Q4 disclosure | Low | High | | S3 GPU debt advance rate | 70% | 55% | 82% | A/B | IREN Blue Owl deal + portfolio estimate | Low | High | | S3 new-money share | 72% | 55% | 85% | E | first-cycle-heavy assumption | Medium (with refi bucket) | Medium | | NVIDIA platform deployment rate | 30% | 20% | 50% | E | scenario assumption on $500B announced | Low (kept out of Sources total) | Medium | **Direction of bias if input overstated:** - Site+shell (L2) conversion overstated → physical formation overstated - Compute (L4) non-hyperscaler share overstated → L4 procurement proxy overstated → sources denominator overstated → intensity understated - Compute (L4) overlap overstated → L4 understated → sources denominator understated → intensity overstated - Customer prepayments (S1) share overstated → S1 dollars overstated → sources overstated → intensity overstated - GPU debt (S3) advance rate overstated → GPU debt (S3) volume overstated → sources overstated - IT-load conversion overstated → all MW-derived layers overstated → physical formation overstated **Highest-sensitivity inputs (top three, ranked by product of range width × central weight):** - L4 non-hyperscaler share and overlap: combined, these define $437B of the Compute (L4) central; a 10-point shift in either moves Compute (L4) by $60-90B - L2 site-formation conversion × $/MW: combined, these define $221B of L2 central; a shift from central to either bound moves L2 by $50-100B - IT-load conversion factor: applies upstream to all MW-derived layers (L1a, L2); a shift from 0.75 to 0.65 or 0.85 moves the MW-derived physical layers by \~10%. Non-MW-derived layers (L3 vendor-dollar backlog, L4 revenue proxy) are not affected by this factor. The Ledger will publish a full sensitivity tornado in Vol 1 (early December 2026) alongside quarterly deployment prints. ## 5\. Three Headline Numbers, in Context Consolidating the analysis so far: **Number 1: Gross third-party financing activity 2026-28** Central estimate approximately $646 billion. Deterministic range: $328B to $1,097B. This measures the total dollar flow through third-party financing structures (debt origination, equity issuance, customer prepayments, vendor credit) attributable to AI infrastructure over the three-year window. It does not measure new physical asset formation, some of which is funded outside these structures. And it includes refinancing activity, which reshuffles existing capital rather than creating new. **Number 2: Third-party-associated formation and procurement** Central estimate approximately $948 billion. Modelled 5th to 95th percentile range under the model's common-factor Monte Carlo assumptions: $681 to $1,523 billion. This measures third-party-associated formation and procurement: new grid capacity, new sites, new equipment, and compute/IT procurement that flows through third-party financing structures rather than being funded directly by hyperscaler balance sheets. It excludes hyperscaler self-funded capex, which is separately tracked at approximately $2.5 to $3.2 trillion over the same window. **Number 3: Refinancing and secondary activity within the stack** Central estimate approximately $156 billion (including S5 secondary equity of $30B central). Range $27 to $583 billion. This measures financing flows that reshuffle existing capital claims rather than funding new formation: GPU debt refinancings, sale-leaseback recapitalisations of existing sites, secondary equity issuance where insiders exit. The lower bound is zero because in a fast-growing new cycle, refinancing volume can be very small; the upper bound is high because if the first-generation debt refinancing window (late 2028) triggers a refi wave, the number could scale considerably. **Reconciliation of the three measurements:** | Measurement | Central | What it measures | Includes | Excludes | | ----------------------------- | ------- | ------------------- | ------------------------------------------------------------ | ----------------------------------------------- | | Gross third-party financing | $646B | Financing flows | Debt, equity, prepay, vendor | Hyperscaler self-funding | | Physical formation (L1-L3) | $511B | New physical assets | Grid access, sites, equipment (excl. GPU/system procurement) | Pure secondary; compute procurement | | Compute & IT procurement (L4) | $437B | Procurement flows | GPU/system procurement proxy | Overlap with hyperscaler-scope activity removed | | Refi / secondary | $156B | Capital rotation | Refi, sale-leaseback, secondary | New formation | **These numbers must not be summed.** They measure different things on different accounting bases. The essay explains why each is analytically distinct in Sections 2-4. The $948B bridge · physical formation + compute procurement The middle headline measurement decomposed into its two diagnostic components. L1-L3 PHYSICAL FORMATION$511Bbottom-up MW × $/MWGrid · sites · equipment · MW-derivedRanges from $215B (low) to $1,027B (high)+L4 COMPUTE PROCUREMENT$437Brevenue-side proxyNVIDIA DC rev × non-hyperscaler − overlapTwo Grade-E inputs; range $316-572B\=COMBINED$948Bformation + procurementDifferent measurement regimes. Read as parallel diagnostic quantities, not as a homogeneous formation figure.L1-L3 is bottom-up MW-derived physical build; L4 is revenue-side procurement proxy. The sum is a summary aggregate. Bottom-up model · both components explicit at central estimate **Adjacent flows for context:** - **Hyperscaler self-funded capex, 2026-28:** approximately $2.5 to $3.2 trillion, of which perhaps 15 to 25 percent ultimately touches assets that also appear in the third-party stack when hyperscalers lease from REITs, purchase compute from neoclouds, or co-invest in projects. - **Upstream model-company equity flowing to compute purchases:** central estimate approximately $64 billion during the window, from total model-company primary and secondary equity of roughly $150 billion of which 30 to 60 percent ultimately flows to infrastructure compute rather than R&D, working capital, or acquisitions. **Framing the numbers correctly:** The AI buildout is majority hyperscaler-financed. The bulk of physical formation over 2026-28 sits on hyperscaler balance sheets; the third-party stack accounts for the remainder. The exact split requires better hyperscaler AI-attributable capex disclosure than currently exists. The remainder is a real market: $0.65 trillion of gross third-party financing activity, funding $0.51 trillion of physical infrastructure formation (L1-L3) and $0.44 trillion of compute and IT procurement (L4) over three years. The $948B formation-and-procurement measurement contains two diagnostic components, each with its own methodology and uncertainty. Reading them as one homogeneous quantity destroys the analytical value of the decomposition. --- ## 6\. The Financing Chain: Three Worked Examples Abstract layer definitions get harder to attack when they are stress-tested against actual deals. Three examples that show the capital-sources architecture in operation. ### Example 1: IREN Blackwell Ultra tranche Physical asset formed: incremental GPU capacity in IREN's Canadian data centre, deployed on Nvidia GB300 hardware. Approximately $2.8 billion of GPU capex including associated equipment and integration. Financing chain, in order of seniority: The transaction spans two distinct financing tranches that must be kept separate. **Non-investment-grade tranche** (Blue Owl-led $2.4 billion financing, with PIMCO serving as adviser to certain investors, at 9 percent, 2.5-year tenor): - Debt: $2.4 billion. The financing package provides approximately 90 percent coverage of associated GPU capex when combined with prepayment support. **The 90 percent is combined financing coverage from debt plus prepayment against the GPU capex denominator. The standalone debt advance rate is materially lower.** - Customer prepayments: 45-55 percent of GPU capex on the same tranche - IREN operator equity: the residual Reconciliation: on a **narrow "GPU purchase price" denominator**, debt plus prepayment exceeds 100 percent of the GPU asset cost. The excess isn't accounting duplication. The financing package generates funding capacity beyond the narrow GPU purchase, with the extra flowing to adjacent site and electrical infrastructure capex for the same deployment. IREN's own commentary confirms this pattern: recent financings plus prepayments can exceed 100 percent of associated GPU capex, with the excess supporting data-centre capex. **Investment-grade tranche** (separate): $3.6 billion at 6 percent coupon supporting Microsoft-anchored capacity, structurally distinct from the Blue Owl financing. Not included in the reconciliation above. Structural read: this is a transaction in which the customer, the lender, the insurance provider (embedded in deal economics), the operator equity holder, and the equipment vendor all take fractional exposure to a single physical asset. Different capital providers underwrite different pieces of the risk. The lender's risk exposure includes GPU residual value, operator credit, customer concentration, and structural protections. The customer underwrites its own ability to consume compute. IREN's equity underwrites operational execution. No single party takes the full stack. ### Example 2: Anthropic-Nscale $45 billion, six-year (illustrative reconstruction, transaction-level financing terms not disclosed) Planned physical deployment: up to 460 megawatts of AI-DC capacity at Nscale's Monarch Compute Campus in West Virginia, running on Vera Rubin. Vera Rubin deployment is expected from late 2027, per press coverage. Announced August 26, 2026. Financing chain, in reverse order (starting from the customer): - Anthropic's compute commitment provides contracted future revenue against which Nscale can finance the deployment over six years. Anthropic is a large, well-capitalised model company with substantial contracted compute demand; press reporting has described the Nscale commitment as being made in the context of Anthropic's ongoing capital-raising cycle. The commitment is upstream demand support that can become financing-enabling (UPSTREAM in the model), not directly infrastructure capital. Precise valuation and IPO-timing claims are deliberately not carried into this financing-chain analysis because they are not required for the mechanism to be illustrated. - Nscale's operator equity funds the balance sheet against which infrastructure debt is issued. Nscale acquired the 2,250-acre Monarch site in March 2026. - The physical build will require a combination of project/equipment financing and operator capital, but the transaction-level financing structure has not yet been disclosed. One possible financing route is a mix of Source 4 (real-estate/project-finance debt) and Source 3 (GPU-collateralised debt), though the specific composition is not observable at announcement. Materially, Microsoft withdrew from an earlier LOI on the same Nscale capacity in the summer, freeing capacity for Anthropic. Two observations: first, Anthropic has announced or reportedly committed to tens of billions of dollars of compute and infrastructure-related capacity in a short period across several distinct structures (compute-purchase agreements, hardware procurement, and cloud contracts). These are not economically equivalent and should not be aggregated into a single "compute-buying velocity" number. Second, Microsoft's withdrawal is an early observable signal that hyperscalers may be becoming more selective about capacity commitments. Whether this is the first evidence of broader capacity discipline is a Ledger question, not an established fact. Both observations will be tracked as the pattern extends or fails to. ### Example 3: Digital Realty to Blackstone, Northern Virginia (reported $7.8 billion gross transaction value) Physical asset formed: none. As reported at announcement, this transaction transfers ownership of 288 megawatts of existing operating data-centre capacity from Digital Realty to a Blackstone-led private capital structure at a reported gross transaction value of $7.8 billion (including assumed debt and remaining capex, per Digital Realty press release). Digital Realty's consideration for Blackstone's interests was approximately $3.5 billion. Announced June 2026. Financing chain: - Blackstone growth-equity vehicles fund the equity portion of the acquisition. - Structured project finance debt funds the debt portion, likely with insurance participation. - Digital Realty receives proceeds, which it will redeploy into new-formation capex. Read: this is Number 3 (refinancing and secondary activity), not Number 2 (physical formation). The 288 megawatts of capacity existed before the transaction and exist after. What changed is the ownership structure and the capital-provider mix. Digital Realty's proceeds will show up in Number 2 when it deploys them into new sites, but the transaction itself is recapitalisation. This is why the three-number decomposition matters. If the essay collapsed all financing activity into one headline, the $7.8 billion here would look like new AI infrastructure capital. It is capital rotating within the stack. --- ## 7\. Risk Transmission: The Matrix The single most useful analytical output of this framework is the transmission matrix: the cross-layer shock-exposure matrix that describes how a shock originating in one layer propagates to others within four quarters. The intuition: a shock in one layer does not stay in that layer. A grid moratorium disproportionately hurts physical uses but eventually reaches financing sources through cancelled projects. A neocloud default hits Layer 5 sponsor equity first, then propagates to Layer 4 GPU debt through collateral impairment and Layer 3 equipment through order cancellation. A model-company valuation shock hits customer prepayment ability first, then propagates through offtake credit doubts to every layer of the stack. The matrix below gives central-case propagation bands representing the approximate share of the target layer economically exposed to an originating shock within four quarters. Bands are directional scenario estimates rather than empirical measurements; they are the framework the Ledger will refine each quarter. The values below use Low / Medium / High / Full bands rather than precise percentages, because these are scenario propagation assumptions rather than empirically calibrated propagation bands. Interpretation: Full = \~100% (the originating layer or its primary exposure); High = \~50-80%; Medium = \~20-50%; Low = \~5-20%; Negligible = <5%. **Important methodological caveat.** The bands represent **exposure** (share of the target category economically connected to the shock cohort), not **loss severity** (percentage of that exposure that ultimately gets impaired). A "Full" band on Neocloud default → S5 means 100 percent of sponsor equity exposure is connected to the affected operators, not that 100 percent of sponsor equity value is lost. A future Ledger iteration will separate exposure × loss severity as two distinct dimensions; the current matrix combines them for scenario-level readability. | Shock | Grid access | Firm supply | Site+shell | Equipment | Compute | Prepay | GPU debt | Sponsor eq | | -------------------------------------- | ----------- | ----------- | ---------- | --------- | ------- | ------ | -------- | ---------- | | Grid constraint / moratorium expansion | Full | High | High | Medium | Low | Neg | Low | Low | | Equipment shortage | Low | Low | Medium | Full | Low | Neg | Medium | Low | | GPU residual value shock | Neg | Neg | Low | Medium | Medium | Low | Full | Medium | | Neocloud default | Neg | Neg | Medium | Medium | Medium | Low | High | Full | | Model-company valuation shock | Neg | Neg | Low | Medium | Medium | High | Medium | High | | Insurance withdrawal | Neg | Neg | Low | Low | Low | Low | High | Low | | Rate shock (+200 bp) | Low | Low | Low | Low | Low | Neg | Medium | Medium | | Regulatory GPU-debt rule change | Neg | Neg | Low | Low | Medium | Low | High | Medium | Transmission matrix · cross-layer shock exposure Bands represent exposure share, not loss severity. Full ≠ 100% loss. Grid accessFirm supplySite+shellEquipmentComputePrepayGPU debtSponsor eqGrid moratoriumFullHighHighMedLowNegLowLowEquipment shortageLowLowMedFullLowNegMedLowGPU residual shockNegNegLowMedMedLowFullMedNeocloud defaultNegNegMedMedMedLowHighFullModel-company markdownNegNegLowMedMedHighMedHighInsurance withdrawalNegNegLowLowLowLowHighLowRate shock (+200bp)LowLowLowLowLowNegMedMedRegulatory GPU-debt ruleNegNegLowLowMedLowHighMedEXPOSURE BANDS · SHARE OF TARGET LAYER ECONOMICALLY EXPOSED WITHIN 4 QUARTERSNeg · <5%Low · 5-20%Med · 20-50%High · 50-80%Full · \~100% Directional scenario estimates. Ledger will refine bands against observed propagation. Reading the matrix: - **Grid constraint** hits Layer 1a hardest by definition. Cascades into Layer 2 and Layer 3 because sites and equipment cannot commission without grid. Cascades into GPU purchases and financing less directly, because GPUs can be redeployed elsewhere and financing can wait. - **GPU residual shock** hits Layer 4 physical uses moderately (writedown, deferred purchases) but hits Source 3 debt hardest because the debt is collateralised against residual value. Cascades into Source 5 sponsor equity through operator valuation impairment. - **Neocloud default** is the most systemic shock in the matrix. Source 5 sponsor equity goes to 100 percent (operator equity is impaired), Source 3 debt to 80 percent (writedown on collateral and refi freeze), and physical uses across Layers 2 through 4 all take material hits through cascaded project cancellation and equipment order deferral. - **Model-company valuation shock** propagates through customer prepayment ability first (Source 1 at 60 percent), then through offtake credit doubts to every downstream layer. - **Insurance withdrawal** is a narrower shock. Its concentrated impact on Source 3 (55 percent) reflects the sensitivity of insurance-enhanced GPU debt to the underlying insurance capacity. - **Regulatory reclassification of GPU debt** operates through Source 3 first (70 percent) as sovereign LP allocations to insurance capacity tighten and the insurance-to-debt multiple compresses. Cascades to Layer 4 physical uses through deferred purchases and to Source 5 through operator equity multiple compression. The matrix is the analytical framework the Ledger will update quarterly. As new shocks materialise (or partially materialise), the propagation bands get refined against observed propagation. The band set becomes progressively more empirical over time. Two immediate uses: **Stress-testing headline numbers.** Applying any single shock's propagation bands to the central-estimate stack values produces a stressed set of layer values. Compound shocks (two or more simultaneously) produce non-additive impacts because propagation assumptions partially overlap. **Identifying binding constraints for the current quarter.** Whichever shock category has the highest observed movement in leading indicators is the shock the Ledger flags as the binding constraint that quarter. Current working hypothesis: grid access is the binding constraint (Abbott ERCOT freeze, FERC's June 2026 show-cause orders reshaping large-load interconnection). This is a research judgement rather than an observed statistic. **Binding constraint definition**: the layer whose marginal availability most limits incremental third-party-associated physical formation during the forecast window. The Ledger will test this through marginal availability and project conversion indicators (queue-to-executable interconnection ratio, project deferral counts, executable capex vs disclosed capex spread) as quarterly data lands. The Ledger will restate this each quarter. --- ## 8\. Capital Concentration: Who Owns the System The mechanical breakdown into five uses and five sources understates what is happening in the stack. The same capital providers appear across multiple sources and multiple structures. That concentration is a defining feature of the current cycle. ### Defining "cross-layer capital provider" "Controlling" overstates the governance role that concentration measures. A more accurate term is **cross-layer capital provider**: a capital participant that holds material positions across three or more distinct stack categories, drawn from the five physical layers, the five financing sources, upstream demand financing, and platform capital. Material is defined as more than $2 billion of deployed capital or more than 10 percent of any single deal. Enforcing the strict three-category threshold produces a smaller list than a loose reading would: Cross-layer capital provider inclusion table (based on publicly disclosed positions as of Aug 2026): **Principal cross-layer capital providers (≥3 stack categories, deployed LP capital):** | Provider | Site | Compute | GPU debt | RE/PF debt | Sponsor eq | Upstream | NVIDIA platform | Stack categories | Qualifies? | | ------------------ | ---- | ------- | -------- | ---------- | ---------- | -------- | --------------- | ---------------- | ---------- | | Blackstone | ✓ | | ✓ | ✓ | ✓ | ✓ | ✓ | 6 | Yes | | MGX (sovereign LP) | | | | ✓ | ✓ | ✓ | | 3 | Yes | | KKR | ✓ | | | ✓ | ✓ | | ✓ | 4 | Yes | | Brookfield | ✓ | | | ✓ | | | ✓ | 3 | Yes | **Strategic/contingent cross-layer participant (separate category):** | Provider | Site | Compute | GPU debt | RE/PF debt | Sponsor eq | Upstream | NVIDIA platform | Stack categories | Role | | -------- | ---- | ------- | -------- | ---------- | ---------- | -------- | --------------- | ---------------- | ---------------------------------- | | NVIDIA | | | | | ✓ | ✓ | ✓ | 3 | Contingent co-investor + strategic | **Material participants below the threshold (Ledger watchlist):** | Provider | Stack categories | Notes | | --------- | ---------------- | ------------------------------------- | | BlackRock | 2 | NVIDIA platform + Site (L2) | | Apollo | 2 | NVIDIA platform + GPU debt (S3) | | Coatue | 2 | Sponsor equity (S5) + Upstream demand | | GSAM | 1 | NVIDIA platform | | PIMCO | 1 | GPU debt (S3) | Cross-layer capital-provider concentration map Documents participation, not exposure. Exposure HHI not yet calculable. SiteComputeGPU debtRE/PF debtSponsor eqUpstreamNVIDIA pltCategoriesBlackstone6MGX3KKR4Brookfield3NVIDIA3BlackRock2Apollo2Coatue2GSAM1PIMCO1Principal (≥3 categories · deployed LP capital)Contingent / strategic (NVIDIA)Watchlist (<3 categories) Public disclosure as of Aug 2026 · investment-manager level Four principal cross-layer capital providers currently meet the ≥3-category threshold with deployed LP capital. NVIDIA is tracked in a separate strategic/contingent category because its exposure structure differs materially: NVIDIA's contingent co-investment sits alongside sponsor LP capital rather than inside the loss-absorbing tranche. Combining NVIDIA with principal capital in a single concentration metric would compare apples to oranges, so the Ledger keeps the two categories distinct. Names that participate materially but do NOT clearly meet the three-layer threshold on current disclosure include BlackRock (platform + Site (L2)), Apollo (platform + GPU debt (S3)), Goldman Sachs Asset Management (platform), Coatue (Sponsor equity (S5) + upstream demand), and PIMCO (Source 3). The Ledger will track their participation quarterly and reclassify as their positioning expands. The pattern: four principal capital providers plus one strategic/contingent participant hold material positions across three or more categories of the stack. Another five participate materially in one or two categories. Whether the boundary count is five, six, or eight is less important than the concentration itself: the AI infrastructure financing stack is intermediated by a small group of cross-layer capital providers whose participation spans otherwise-independent layers. Whether that participation translates into correlated exposure is a downstream research question the Ledger tracks separately. **Participation concentration is distinct from exposure concentration.** The Concentration Map above documents cross-layer *participation*, meaning which providers appear across how many categories, and this is observable from public disclosure. It does not document cross-layer *exposure*: how much aggregate capital each provider has deployed across those categories, weighted by loss severity. Exposure concentration is not yet observable at portfolio level from public disclosure. F4 (below) uses a pre-specified analytical threshold for the top-four exposure share; the Ledger will replace it with a formal exposure HHI or top-four share once the exposure denominator becomes calculable from disclosure or from fund-level bottom-up tracking. ### Three implications **Diligence-compression flywheel.** Repeated participation across the same institutions can reduce information asymmetry and shorten syndication timelines because counterparties, documentation and operating assumptions become familiar. Prior underwriting experience within the network can therefore compress subsequent deal timelines. Information-sharing across concentrated names compresses the diligence timeline for each subsequent deal. Capital velocity can rise as a result. **Correlated repricing risk.** If any single one of these names pulls back, the effect propagates through multiple layers at once. Blackstone reducing its AI-infra exposure by 20 percent would affect the NVIDIA platform, Source 5 growth equity, Source 4 real estate, and Layer 2 REIT-adjacent structures in the same quarter. The layers are structurally independent. The capital providers are not. **Common-risk-factor exposure.** The AI infrastructure stack is not a CLO market. But it has the beginnings of a common-risk-factor problem. The same capital providers are simultaneously exposed to the same underlying AI demand shock through multiple legal entities and financing structures. If AI demand disappoints, correlated marks and risk limits can propagate across multiple structures these names participate in, though contractual seniority and loss profiles will differ by exposure and prevent uniform simultaneous writedowns. That is different from a pure asset-correlation problem (it is a capital-provider-correlation problem. As a working hypothesis: the AI infrastructure stack may be less asset-correlated than the 2007 financial system but more capital-provider-correlated. That framing is measurable, and the Ledger will track it. ## 9\. The NVIDIA Position Section 8 identifies NVIDIA as a strategic-contingent participant separate from the four principal cross-layer capital providers. That treatment understates something specific about NVIDIA's role that no other participant in the stack shares. NVIDIA sits simultaneously in three distinct relationships with its own customers. - **Revenue vector.** NVIDIA earns its data-centre revenue (roughly $400 to $550 billion in 2026 depending on the run rate assumed) when neoclouds, model companies, and hyperscalers buy accelerators and associated systems. - **Financing vector.** NVIDIA holds a contingent co-investor role in the six-sponsor platform announced August 10 2026 alongside Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR at **$500 billion of announced capacity** (not deployed capital). The platform is designed to provide sponsor and debt capital to AI infrastructure operators, some of which are also NVIDIA customers. The degree to which deployed platform capital ultimately finances NVIDIA purchases is not yet observable at portfolio level; the essay's central assumption is a 20-50 percent conversion of announced capacity, Grade E. - **Strategic-support vector.** Direct credit, project-level commitments, and ecosystem investments to specific buyers and their infrastructure. Reported examples during 2026 include a guarantee of up to $105 billion supporting an OpenAI-anchored Ohio data-centre lease, alongside NVIDIA's $1.5 billion investment in SB Energy; ecosystem equity in CoreWeave; direct participation in model-company rounds; and a $3.5 billion MediaTek convertible-bond commitment. A guarantee is not equivalent to a $105 billion NVIDIA financing outlay, but it is a contingent balance-sheet exposure with a similar systemic transmission if triggered. The exact aggregate is not disclosed at portfolio level. The NVIDIA position Hardware sales, platform financing, and project guarantees to overlapping counterparties. NVIDIA3 vectors↺ THE LOOP · FINANCED BUYERS BECOME REPEAT PURCHASERSREVENUE GPUs and systems sold to neoclouds, model companies, and hyperscalers. NVIDIA data-centre revenue $400-550B/yr (2026 range) FINANCING Contingent co-investor in the 6-sponsor platform ($500B announced capacity, not deployed) that provides sponsor and debt capital to AI infrastructure operators. Apollo · BlackRock · Blackstone · Brookfield · Goldman Sachs · KKR STRATEGIC SUPPORT Direct credit, project-level guarantees, and ecosystem equity in specific buyers and their infrastructure. $105B guarantee on OpenAI Ohio · $1.5B SB Energy · $3.5B MediaTek · CoreWeave Revenue and repeat purchases loop back into NVIDIA data-centre revenue growthSIMILAR TO CISCO 2001 · STRUCTURED DIFFERENTLYCisco 2001 (telecom cycle) - Vendor extended equipment financing directly - \~$2.7B loan-and-lease provisions absorbed on Cisco balance sheet when cycle turned - First-loss concentrated on one vendor NVIDIA 2026 (AI cycle) - Contingent co-investor via 6-sponsor platform - First-loss distributed across sponsors, insurers, lenders - Same demand assumption; wider distribution of exposure Structural read of the six-sponsor platform + reported project commitments. Full deep-dive in F3. **Why this differs from a normal component vendor.** Applied Materials sells lithography equipment to TSMC. TSMC does not rely on Applied Materials capital to buy it. Micron sells memory to hyperscalers. Hyperscalers do not depend on Micron capital to fund the purchases. NVIDIA's revenue and NVIDIA's ecosystem-financing capacity are ultimately exposed to the same underlying AI-compute-demand assumption, even though they are not economically equivalent exposures. If that assumption weakens, all three vectors can move together. ### The Cisco 2001 comparison, more precisely Cisco extended equipment financing directly to customers during the late-1990s telecom buildout, taking equity and debt positions in exchange for continued equipment purchases. When the cycle turned in 2001, Cisco disclosed approximately $2.7 billion of loan-and-lease provisions absorbed on its own balance sheet per its 2001 annual report and contemporaneous SEC filings. The NVIDIA structure is materially different in two ways. First, the six-sponsor platform means principal capital sits with the sponsors and their LPs; NVIDIA participates as a contingent minority co-investor. First loss on any specific transaction reaches Blackstone, KKR, Brookfield or Apollo funds before it reaches NVIDIA's balance sheet. Second, insurance-enhanced structures from Aon, Marsh and reinsurance markets absorb specified project risks that were entirely uninsured in the 2001 vendor-financing pattern. That distribution improves resilience against any single-name failure. It does not remove ecosystem-level correlation. If GPU residual values compress in a cycle turn, NVIDIA revenue slows, the platform's contingent capital capacity tightens, and NVIDIA's balance-sheet strategic commitments come under stress in the same quarter. The concentration is on the demand-side assumption, not on any single vendor balance sheet. ### When the concern actually activates Correlation here matters more than any single vector. In a normal cycle the three vectors are only loosely correlated: revenue depends on end-demand, ecosystem financing depends on capital-market conditions, and strategic commitments depend on specific relationship management. In a stressed cycle they can converge on the same latent variable: realised AI compute demand. Call this **state-dependent ecosystem circularity**: the correlation exists but activates asymmetrically under stress. **Market sensitivity is already visible.** On August 27 2026, NVIDIA reportedly paused a recently launched revenue-sharing financing initiative for small AI-cloud companies amid investor concerns about circularity and ecosystem control (per Reuters). This is evidence of market sensitivity to the mechanism, not evidence that the mechanism has generated losses. The distinction matters: the concern here is not that circular losses have materialised but that the ecosystem structure has become the object of active investor scrutiny. That is the amplifier §13 identifies. The isolated financing cascade in that section assumes the shock stays inside the financing system. The NVIDIA position means a shock that starts in end-demand hits revenue, ecosystem-financing capacity, and strategic commitments simultaneously and in the same direction. ### What the Ledger will track The recurring Ledger monitors six NVIDIA-specific observables under Section F: - Platform deployment rate: contracted transactions as a share of the $500 billion announced capacity. - Any disclosed writedown on ecosystem-related receivables, contingent commitments, or platform capital. - Concentration of NVIDIA-financed operators as a share of total neocloud debt outstanding. - Any pullback in NVIDIA's direct project-level commitments (subsequent transactions on the OpenAI/Ohio scale or absence of them). - Second-derivative signals: NVIDIA guidance changes, capex commentary, or margin-mix disclosure that would indicate demand-side pressure. - **Revenue-financing linkage.** Share of NVIDIA data-centre revenue attributable to customers receiving NVIDIA-supported financing, directly or indirectly. This is the killer circularity KPI. NVIDIA does not currently disclose it. The Ledger records the disclosure gap itself as a monitored condition: if the intersection remains unmeasurable through 2027, the systemic-transparency assumption behind the current benign read of the structure is not testable. F3, publishing later in the series, treats the six-sponsor platform in depth as a category of financing structure with NVIDIA's role as the case study. F5 treats the model-company layer, including how model-company equity capitalises the same demand curve NVIDIA depends on. This section identifies the position; those later essays quantify it. --- ## 10\. Duration Mismatch: The Systemic-Risk Chapter Duration matters more than any dollar figure in this stack. The asset life of a modern AI data centre is 10 to 20 years for the shell, 5 to 10 years for the electrical equipment, and 3 to 6 years for GPU economic collateral life (the period over which GPU residual value supports the debt structure that funded it). The customer offtake contract runs 3 to 7 years. The debt runs 2 to 4 years. The sponsor equity is 5 to 7 years for private, indefinite for listed. The model company's own equity funding cycle is 12 to 24 months between rounds. The stack generally contains shorter-duration claims layered on longer-lived assets. In the structures examined here, debt is typically shorter than offtake, offtake shorter than GPU economic collateral life, and GPU economic life shorter than site life. Grid access is governed by a different contractual and regulatory duration structure that resists direct comparison with the other layers. Duration mismatch · why refinancing is the systemic-risk point Ranges observed in currently-disclosed structures. Grid access resists direct comparison. 0yr5yr10yr15yr20yrSite / shell life10-20yrElectrical equipment life5-10yrGPU economic collateral life3-6yrCustomer offtake tenor3-7yrSponsor equity (private) tenor5-7yrDebt tenor2-4yrModel-company equity cycle1.0-2yr↑ SHORTER-DURATION CLAIMS LAYERED ON LONGER-LIVED ASSETSDebt written in 2026 matures in 2028-29 · GPU is \~1 generation behind by then · refinancing must clear against aging collateral. Structures examined here. Extending offtake tenor beyond 7 years is a durability signal. That mismatch is the systemic risk in the stack. 2028-29 refinancing cliff is a systemic risk Duration mismatch anchored at 2026 origination. Shorter-duration claims on longer-lived assets. 2026202720282029203020312032Debt (2.5-yr Blue Owl print)Refi window opens · lender must show upCustomer offtake (3-7 yr)Renewal / repricing risk at 3-5 yrGPU economic collateral lifeBy 2028-29 the chip is roughly 1 generation behindSite + shell lifeOutlives everything above (but capped by chart span)▼ THE 2028-29 REFI CLIFFWhy the cliff matters.When 2026-origination debt matures at 2028-29, the GPU collateral is roughly one generation behind and the offtake has 6-24 months of tenor remaining. Refi terms depend on the next lender showing up. If they do not, the cascade in §12 is fast. Structures examined in this essay · durations are Ledger-tracked as they evolve The mechanism: a GPU deployment funded by 2.5-year debt against a 5-year offtake against a 5-year GPU economic collateral life requires that when the debt matures at year 2.5, a new lender is willing to refinance the remaining collateral, which is by then technologically 2.5 years old, or roughly one generation behind the leading deployed architecture depending on shipment cadence. The refi terms depend on the new lender's view of residual value, on the operator's remaining offtake tenor, on the current cost of capital, and on whether the technology curve has continued to compound as expected. If any of those inputs deteriorates, refinancing pricing widens, refinancing volumes compress, or refinancing may not clear at all. And the operator does not have the option of holding to maturity (the debt matures and the payment must be made from either refinancing proceeds or from operating cash flow. The current architecture is durable only if capital can be repeatedly rolled, repaid, or replaced across technology generations. That is a stronger condition than saying "the GPU market will grow." It requires that at every point in the debt roll cycle, there is a lender willing to underwrite next-generation collateral against remaining-tenor offtake. If that lender exists reliably, the stack is durable. If that lender disappears for even one refi wave, the stack has a cascade problem. Watch the refi spread. Deployment growth lags. The spread moves first. If the next IREN-scale deal prices at 10-11 percent against Blue Owl's original 9, that's amber. A deal that fails to refinance at all is red. When a lender publicly walks away from AI-infra credit, you're past the individual-deal question. That's when the system is moving. Refinancing wave visibility · illustrative scenario First-generation 2.5-year GPU debt originated 2026 refinances mid-late 2028\. This is the systemic pressure point. $0B$20B$40B$60BQ3 26Q4 26Q1 27Q2 27Q3 27Q4 27Q1 28Q2 28Q3 28Q4 28Q1 29Q2 29First-gen refi window opensQ3 2028: Blue Owl-IREN tranche + comparable 2.5-yr debt reprice Illustrative timing. Ledger's H.4 refi-window visibility metric tracks actual maturity ladder. The GPU financeability equation, as a framework: **The three-condition financeability test: cash-flow durability, collateral durability, refinancing availability.** A conceptual test, not a calibrated quantitative score. All three conditions must hold. Financeability approaches zero if any single condition approaches zero. - **Cash-flow durability** depends on: offtake tenor, renewal probability, termination protection, counterparty credit, customer concentration, and contract enforceability. A 3-year contract with 95 percent renewal probability and strong take-or-pay is not necessarily worse than a 7-year contract with weak termination protection. A 5-year contract with 95 percent renewal probability is materially weaker if 80 percent of operator revenue comes from one model company. - **Collateral durability** depends on: the GPU residual value curve and secondary-market liquidity. - **Refinancing availability** depends on: lender depth, spread environment, and regulatory treatment. Any one of these three going to zero brings financeability to zero. All three are independently variable. Each requires its own monitoring metric in the Ledger. The reason this matters more than the dollar quantities: the dollar quantities describe the current state. The duration mismatch describes what happens as the current state ages. Debt written in 2026 matures in 2028-29\. If refinancing conditions in 2028-29 are different from origination conditions in 2026, the stack reprices mid-cycle rather than at end-of-cycle. Mid-cycle repricing is where the systemic pain lives. --- ## 11\. Historical Parallels Three structural precedents deserve comparison. ### 2001 Cisco vendor financing Cisco extended equipment financing to customers during the late-1990s telecom buildout. When the cycle turned in 2001, Cisco absorbed approximately $2.7 billion of loan-and-lease provisions on its own balance sheet (per Cisco's 2001 annual report and contemporaneous SEC filings). The NVIDIA position (§9) rhymes with that pattern with a structural difference in first-loss distribution. The read: the current structure spreads exposure across sponsors, insurers, and LPs rather than concentrating it on one vendor balance sheet, which improves resilience against any single-name failure but does not eliminate ecosystem-level correlation to the underlying AI-demand assumption. §9 develops this position in detail. ### 2007 CLO cycle Collateralised loan obligations pre-2008 had two structural features that combined to produce correlated repricing. First, concentration of tranche investors held positions across many CLOs simultaneously. Second, the underlying loans were originated to below-investment-grade corporate borrowers whose creditworthiness deteriorated in a correlated way during the cycle downturn. The current AI-infra financing stack matches the first feature. The cross-layer capital-provider list above shows the pattern. It differs on the second feature: the underlying credit is contracted offtake from hyperscalers (investment grade) and from model companies whose equity value depends on end-customer demand for AI services, rather than below-investment-grade corporate paper. CLOs blew up because the underlying loans moved together. Same cycle, same borrowers, same time. The AI infrastructure stack behaves differently. Grid access, sites, GPUs, and model-company equity respond to different inputs. Structurally independent assets. But the capital providers behind them are the same handful of firms. Blackstone underwrites Layer 2 real estate. Blackstone Growth writes Source 5 sponsor equity. Blackstone Tactical Opportunities holds upstream demand exposure. Add MGX, KKR, Brookfield and NVIDIA to the same names, and four otherwise-independent layers reprice whenever any one of them rebalances portfolio risk. The mechanism is unlike CLOs. The outcome rhymes. While hyperscaler offtake holds and AI demand keeps compounding, this correlation is manageable. When either fails, the same names pull back everywhere at once. ### Project finance for pipelines and oil sands Long-duration project finance for hydrocarbon infrastructure, particularly the 2005-2010 Canadian oil sands and North American pipeline cycles, is the closest structural analogue for durability. Project finance combined long-term offtake contracts with investment-grade counterparties, insurance-supported debt structures, layered sponsor equity, and standardised documentation across deals. These structures demonstrated that long-duration contracted cash flows can support durable financing across commodity cycles. The current stack matches the layered structure. It matches the insurance-supported debt with sponsor co-investment. It matches the discipline of contracted offtake as the primary credit driver. It does not yet match the offtake tenor. Project finance offtake ran 20 to 25 years; AI-infra offtake currently runs 3 to 7 years. The read: the stack is a hybrid. Project-finance-like structure with growth-equity-like underlying credit and short-duration technology cycles. Durability depends primarily on whether offtake tenors extend as the market matures. Any Nscale, IREN, CoreWeave, or platform-structured deal that materially extends offtake tenor beyond 7 years is a positive signal for stack durability. Any that shortens the tenor, or fails to close at expected pricing, is a negative signal. --- ## 12\. Stress Tests and Falsifiers A methodological note before the falsifiers. The transmission matrix in the previous section reports **stock impairment**: the fraction of a target layer's central value that gets impaired by an originating shock. Stock impairment is not identical to flow reduction. A 30 percent collateral value impairment on GPU debt does not automatically produce a 30 percent reduction in physical GPU deployment. It produces refinancing spread widening, tighter LTV requirements, and equity-cheque-size increases on new deals, all of which reduce new-project origination by a smaller percentage than the underlying impairment. **Stock-to-flow response factor: 0.3 to 0.6.** Evidence grade E (analyst scenario assumption). This is a lender/originator behavioural-response assumption, not a mechanical accounting conversion. It reflects the general pattern that impairments in a layer's stock value translate to a smaller percentage reduction in new-flow origination because pipeline projects continue, existing offtake pays, and lenders reprice rather than exit. The specific response mechanisms (advance-rate haircuts, tenor shortening, pricing widening, additional equity requirements, tighter offtake requirements, origination pauses, market exits) each map differently from stock impairment to flow reduction; the 0.3-0.6 range is a scenario-level average across those channels. The Ledger will refine this response factor as new evidence lands. The dollar estimates below apply this conversion and each represents a **scenario estimate** rather than a mechanically-derived model output. Ranges assume 0.3-0.6 stock-to-flow conversion combined with market-wide repricing at severity levels stated in each case. **Stress tests vs falsifiers.** The scenarios below combine two distinct categories: - **Stress tests** evaluate whether the architecture holds under adverse conditions. Neocloud default, insurance withdrawal, Series H markdown, grid moratorium expansion, rate shocks: each of these stresses the architecture but does not necessarily prove its central proposition wrong. - **Falsifiers** are evidence that would prove the architecture itself wrong. A short list of thesis falsifiers appears at the end of this section, distinct from the stress scenarios above. The stack has a defined set of shocks that would compress the aggregate distribution materially. Applying the transmission matrix produces quantitative propagation estimates for each. In severity order: **Correlated 2+ neocloud defaults in a single quarter.** Two operators missing debt covenants or defaulting simultaneously would trigger: - Source 3 GPU debt: -80% of central at affected operators, cascading to -30% market-wide as spreads reprice - Source 5 sponsor equity: -100% at affected operators, -25% market-wide as multiples compress - Layer 4 GPU capex: -45% at affected operators through cancelled orders, -15% market-wide Illustrative impact bridge: - Two $5B+ operators default. Each operator has approximately $5B outstanding debt plus $3B associated equity plus $12B committed but undrawn capex, giving a per-operator affected base of $20B and a combined point-estimate of $40B. The $40-60B range widens the upper bound to account for uncertainty in the committed-capex figure (which can be higher for large operators with active expansion programmes) - Direct debt impairment: Source 3 -80% at the two affected operators × 0.5 stock-to-flow = -$4-6B in new debt origination reduced at affected operators (already-drawn debt is a stock impairment, not a flow reduction) - Cascaded flow reduction (not additive to the direct impairment): market-wide spread widening \~150-300bp on remaining Source 3 pipeline reduces new debt origination by $25-45B across the market - Cascaded Source 5 sponsor equity: multiple compression 30-50% at surviving operators reduces new equity origination by $12-30B - Cascaded L2/L3/L4 physical formation reduction: not summed independently of the financing reductions above, but arising from them via the transmission chain; estimated at $50-100B on a scenario basis **Note on double propagation.** The physical-formation impact is a downstream consequence of the financing shock and is not summed independently of the financing reductions. The aggregate impact range synthesises the joint financing and formation channels rather than adding them. Illustrative aggregate range: -$100 to -$180 billion on gross financing activity (scenario synthesis of direct impairment plus market-wide spread widening, not arithmetic sum of preceding channels), -$60 to -$130 billion on physical formation as a downstream consequence. Observable: any $5B+ operator missing a payment or violating a covenant. **Insurance-reinsurance withdrawal.** Removal of a material fraction of AI-DC lifecycle insurance capacity would: - Directly reduce the addressable capex base for insurance-supported debt structures. The dollar impact depends on the specific leverage mechanism between insurance capacity and incremental debt capacity, which is not currently disclosed by lenders at deal level. - Cascade to Source 5 through operator equity multiple compression on debt-constrained operators. - Cascade to Layer 4 through deferred GPU purchases at operators that cannot substitute alternative debt at similar pricing. Illustrative impact: primarily affects highly insurance-dependent GPU financing structures. Precise dollar impact requires a calibrated insurance-to-debt-capacity mechanism that will be built into the recurring Ledger as more disclosure lands. Observable: Q1 2027 program renewals and any material reinsurer withdrawal announcement. **Model-company Series H mark-down of 50% or more.** Anthropic or OpenAI Series H investors marking their positions down materially: - Source 1 customer prepayments: -60% of central through weakened offtake credibility - Source 3 GPU debt: -40% through offtake counterparty credit doubts - Source 5 sponsor equity: -55% through end-demand doubts - Cascades to Layers 2 through 4 through project deferrals Aggregate impact: -$200 to -$400 billion on gross financing, -$150 to -$300 billion on physical formation. Observable: quarterly LP letters and secondary market pricing on Anthropic and OpenAI. **Regulatory treatment change on GPU-collateralised paper.** More realistic than a single SEC/FSOC reclassification event, this covers bank capital treatment changes, asset-concentration limits, insurance treatment revisions, private-credit disclosure requirements, or accounting treatment changes that materially increase capital charges or restrict institutional allocation. Any of these would tighten sovereign LP allocations and insurance-to-debt multiples: - Source 3 GPU debt: -70% through sovereign LP allocation tightening and insurance-to-debt multiple compression - Source 5 sponsor equity: -35% at debt-constrained operators - Layer 4 physical uses: -25% through deferred purchases Aggregate impact: -$100 to -$200 billion on gross financing, similar magnitude on physical formation. Binary risk. Currently green. Observable: any FSOC or SEC-issued guidance addressing GPU-collateralised debt specifically. **Grid moratorium expansion.** If two additional states beyond Texas impose data-centre interconnection moratoria: - Layer 1a: -30% at affected jurisdictions - Cascades to Layer 2 through cancelled site construction - Cascades to Layer 3 through order deferrals Aggregate impact: -$70 to -$140 billion on physical formation, -$50 to -$100 billion on financing. Observable: state-level bills in Texas, Virginia, Ohio. **US-China escalation forcing export-control writedown.** Significant tightening of export controls requiring writedown of installed GPU base. Potentially material. Not currently quantified due to insufficient disclosure on the size of the affected installed base and applicable writedown percentage. The Ledger will build a calibrated estimate as export-control policy evolves. Any single stress scenario compresses the central estimate by roughly 10 to 30 percent, depending on the shock. Two occurring simultaneously produce non-additive impacts because scenario propagation assumptions partially overlap. Three or more push the stack into a different distribution regime that the current model does not fully price. ### Thesis falsifiers Distinct from the stress scenarios above. The scenarios test whether the architecture survives adverse conditions. The falsifiers below test whether the architecture is right in the first place. - **F1\. Third-party financing fails to become material.** If the External New-Money Coverage Ratio remains structurally below approximately 20 percent through 2028 despite continued physical formation, the "financing threshold" thesis is overstated (analytical threshold, not empirical benchmark; used to define the falsifier rather than reflecting a historical cutoff). The buildout is still hyperscaler-dominated, not multi-layer. - **F2\. Financing remains bundled rather than separated by claim.** If GPU, site, offtake, and operator risk continue to be co-underwritten on the same balance sheet rather than independently priced and separately underwritten by distinct capital providers, the risk-separation thesis fails. Operational test: do deal disclosures identify separately priced claims on distinct risk pools, or do they show a single lender/sponsor absorbing bundled exposure? - **F3\. GPU collateral never becomes financeable independently of hyperscaler credit.** If lenders require hyperscaler guarantees for essentially all GPU-backed debt, the Blue Owl-IREN print proves unrepresentative and the asset-level financing thesis is overstated. - **F4\. Cross-layer capital-provider concentration fails to persist or reverses.** If the share of identified cross-layer exposure held by the top four principal providers falls below approximately 40 percent (analytical threshold, not empirical benchmark; the current aggregate exposure share is not yet calculable from public disclosure), or if new independent lenders and sponsors materially broaden the provider set, the capital-provider-correlation argument weakens. The Ledger will eventually replace the 40% threshold with a formal cross-layer exposure HHI or top-four share once the exposure denominator is calculable. - **F5\. The 2028-29 refinancing window clears without stress.** If first-generation debt refinances at stable or tighter spreads despite aging GPU collateral, the duration-mismatch chapter overstates the systemic mechanism. - **F6\. Risk transfer does not occur despite financing separation.** If the External New-Money Coverage Ratio rises materially but the Risk Transfer Ratio (H.6) and Recourse-Adjusted Financing (H.7) metrics remain low because hyperscaler guarantees, minimum-purchase commitments, cross-recourse, and sponsor support continue to bear most economic downside, then the financing-modularity thesis holds but the economic-risk-separation thesis fails. This is a distinct falsification: the architecture would then be legally modular but economically coupled to hyperscaler balance sheets. Each falsifier has a defined observable. F1 tracked quarterly via the External New-Money Coverage Ratio. F2 via deal-structure disclosures in Section D of the Ledger. F3 via each new GPU debt deal added to the Blue Owl-IREN benchmark curve. F4 via the Cross-Layer Capital-Provider Map. F5 via debt pricing prints as the refinancing window approaches. Stress scenarios and thesis falsifiers are tracked separately in the Ledger, each with green/amber/red status. ## 13\. Where This Breaks Section 12 catalogues stress scenarios and thesis falsifiers separately. The failure mode this framework worries about most is neither a single stress event nor a single falsification. It is the compound cascade: one triggering event propagates through the stack over 4-8 quarters while capital-provider concentration and NVIDIA circularity amplify each stage. ### The canonical cascade A single $5B+ neocloud misses a debt covenant. Trace what happens over the next 4-8 quarters through the transmission matrix in §7: - **Stage 1 · Direct impairment (quarter of the miss).** S3 GPU debt at the affected operator marks down 80 percent. Direct impact: -$4 to -$6B in new debt origination reduced at that operator. Already-drawn debt is stock impairment; new flow is what compresses. - **Stage 2 · Market repricing (Q+1 to Q+2).** GPU debt (S3) spreads widen 150 to 300 basis points market-wide as lenders reassess GPU collateral value. New debt origination across the rest of the market falls -$25 to -$45B. - **Stage 3 · Sponsor equity compression (Q+2 to Q+3).** Multiple compression 30 to 50 percent at surviving operators. Sponsor equity (S5) new equity origination falls -$12 to -$30B. - **Stage 4 · Refi failure (Q+3 to Q+4).** Any tranche coming up for refi finds the market pricing widely different terms. Some tranches do not clear at all. This is the moment §10's systemic-risk chapter identifies. - **Stage 5 · Formation contraction (Q+4 to Q+8).** L2 site development pauses. L3 vendor orders defer. L4 GPU procurement compresses. Cascaded physical formation reduction -$50 to -$100B. Aggregate at the end of eight quarters: **\-$100 to -$180B on gross financing** and **\-$60 to -$130B on physical formation** as a downstream consequence. This is a scenario synthesis, not an arithmetic sum of the preceding channels; the physical-formation impact is a knock-on of the financing shock rather than an independent addition. Fail cascade · one neocloud default over 4-8 quarters Central column: propagation through the transmission matrix. Side boxes: what makes an isolated shock systemic. TRIGGER · Q0Single $5B+ neocloud misses a debt covenantStage 1 · Q0S3 direct impairment: -$4 to -$6BStage 2 · Q+1 to Q+2S3 spreads widen 150-300bp: -$25 to -$45BStage 3 · Q+2 to Q+3S5 sponsor equity compression: -$12 to -$30BStage 4 · Q+3 to Q+4Refi failure at the aged-collateral tranchesStage 5 · Q+4 to Q+8L2/L3/L4 formation contraction: -$50 to -$100BAGGREGATE · Q+8\-$100 to -$180B gross financing\-$60 to -$130B physical formationAMPLIFIER 1Provider concentration 4 principals + NVIDIA span 3+ categories each. A 20% pullback by any one hits multiple layers in the same quarter. AMPLIFIER 2NVIDIA circularity GPU residual compression hits NVIDIA revenue, its platform capital capacity, and end-market demand at the same time. Cisco 2001 pattern with wider distribution. Scenario synthesis. Ledger tracks leading indicators for each stage under Section I. ### What makes this systemic The isolated cascade above is manageable. Two features make it systemic. **Capital-provider concentration.** The same names hold positions across debt, equity, real estate, upstream demand, and the NVIDIA platform. Blackstone Growth (S5), Blackstone Real Estate (S4), Blackstone Tactical Opportunities (upstream), and Blackstone-adjacent NVIDIA platform positions: if Blackstone reduces AI-infra exposure by 20 percent for portfolio-risk reasons, four otherwise-independent layers reprice in the same quarter. This pullback is not the base case. The relevant fact is that the concentration exists and the transmission mechanism is real. If any of Blackstone, MGX, KKR, Brookfield, or NVIDIA pulls back, the correlation propagates faster than an asset-diversified model would suggest. **NVIDIA ecosystem circularity.** The three-vector NVIDIA position (revenue + ecosystem financing + strategic support) documented in §9 amplifies the cascade because all three vectors capitalise the same underlying AI-demand assumption. If GPU residual values compress, NVIDIA revenue slows, the six-sponsor platform's contingent capacity tightens, and NVIDIA's direct commitments come under stress in the same quarter. §9 explains the position and its Cisco 2001 comparison in detail; the amplifier here is that a demand-side shock reaches all three vectors simultaneously. ### What tells you the cascade is happening Each of the following is a leading rather than lagging indicator. The Ledger tracks each as green/amber/red per Section I. - Refi spread on any new GPU-collateralised deal above 10 to 11 percent, up from the Blue Owl-IREN 9 percent print. - Any $5B+ operator missing a payment or violating a debt covenant. - Aon or comparable insurance capacity contracting quarter-over-quarter, down from $5B in April 2026. - Any of Blackstone, MGX, KKR, or Brookfield publicly reducing AI-infrastructure allocation. - NVIDIA writedown on ecosystem-related receivables, contingent commitments, or platform capital. - Series H mark-to-market on Anthropic or OpenAI compressing 20 percent or more. None of these are currently flashing red. Some are amber. The value of naming the observables is that they can be tracked as leading signals rather than reconstructed after the fact. --- ## 14\. The Recurring Ledger The framework above becomes a recurring quarterly column: the AI Infrastructure Financing Ledger. Each Ledger volume updates the model with new data, refreshes the transmission matrix propagation bands against observed events, and tracks changes in capital concentration, refinancing spread, and layer-specific sensitivities. The Ledger format: **Section A. Three headline numbers, updated.** Uses total, sources total, refinancing total. Change from prior quarter and year-ago quarter noted. Correlated Monte Carlo re-run. **Section B. Layer-by-layer contribution.** Each layer's central estimate, share of total, and quarterly change. Deltas explained. **Section C. Sources decomposition.** Customer prepayments, vendor finance, GPU debt (new money vs refi), RE/PF debt, sponsor equity, insurance-enhanced share, NVIDIA-platform share. Each source's quarterly flow and market pricing point. **Section D. Debt pricing prints.** Every new GPU-collateralised or neocloud debt deal added to the Blue Owl-IREN benchmark curve. Coupon, tenor, insurance participation, advance rate, spread over comparable investment grade, customer prepayment share. **Section E. Insurance capacity movements.** Aon, Marsh, Willis, reinsurance participants. Aggregate capacity, program-by-program, coverage type mix. **Section F. Cross-layer capital concentration.** Updated cross-layer capital-provider map. Any material change in cross-layer participation. **Section G. Refinancing wave visibility.** Debt-maturity schedule across the stack. Which deals refinance in the next 2 quarters, at what expected pricing, with what refi risk. **Section H. Capital-stack diagnostics.** - **H.1\. External New-Money Coverage Ratio.** Identified external new-money sources divided by third-party-associated formation and procurement, decomposed by category (debt / customer-funded / sponsor / vendor). Measures identified capacity-side coverage of the aggregate use. "Coverage" (not "intensity" or "penetration") because both of the latter imply a deal-level allocation between sources and uses that public disclosure does not currently support; the Ledger will migrate to a true penetration measure when the deal-level mapping is available. - **H.2\. New-money share of gross financing.** New-money financing divided by total third-party financing activity. Measures how much of the financing flow represents net new capital versus refinancing recirculation. Rising values indicate cycle expansion; falling values indicate maturity as refinancing share grows. - **H.3\. Funding coverage.** For each major operator: contracted capex against customer prepayment plus committed debt plus sponsor equity. Ratios above 100 percent signal funding capacity beyond the narrow asset denominator; ratios below 100 percent indicate greater dependence on additional funding sources. - **H.4\. Refinancing window visibility.** Debt maturity ladder by operator, GPU vintage, offtake remaining, LTV, refinancing risk score. - **H.5\. Duration mismatch tracking.** Weighted average asset life, offtake tenor, debt tenor, sponsor equity tenor by layer. Changes noted. The Risk Transfer Test · what actually leaves the hyperscaler balance sheet The Ledger's signature test. Legal separation is easy; economic risk transfer is hard. The gap is where the Ledger measures. STAGE 1Nominally external financing Legal separation moved the asset off the hyperscaler balance sheet. $0.65T the reported gross financing figure STAGE 2Recourse-adjusted (H.7) Subtract the portion secured by take-or-pay, minimum-purchase commitments, cross-recourse and vendor guarantees. ? measured quarterly by H.7 · baseline in Ledger Vol 1 STAGE 3Economic exposure Further subtract strategic dependence: capacity the hyperscaler realistically has to preserve to keep the model-serving product functional. ? hidden exposure the balance sheet does not carry as a line item STAGE 4True risk transfer (H.6) The residual: capital genuinely at risk on capital-provider balance sheets rather than economically bearing back to the hyperscaler. ? measured quarterly by H.6 · the number that matters for underwriting H.6 Risk Transfer Ratio + H.7 Recourse-Adjusted Financing · baselined in Ledger Vol 1 (December 2026) - **H.6\. Risk Transfer Ratio.** Share of third-party-associated formation and procurement for which a material portion of economic downside is contractually borne outside the hyperscaler/operator balance sheet. Denominator is the aggregate $948B formation + procurement figure (not the $511B L1-L3 subset) because H.6 is intended as the eventual economic-risk-transfer metric for the whole financing stack, not for the site/grid/equipment layers alone. Component contributions: non-recourse GPU debt (transfers creditor exposure subject to collateral), insurance (transfers specified insured risk), sponsor equity absorbing first losses (transfers first-loss risk). Customer take-or-pay is **not** automatically counted as risk transfer because it can keep material economic risk with the hyperscaler through minimum-purchase obligations, termination liabilities, cross-recourse, and strategic dependence. Distinct from the Coverage Ratio: measures economic risk-bearing outside operator, not nominal financing flow. Full definition and initial measurement in Ledger Vol 1. - **H.7\. Recourse-adjusted financing.** Gross external financing risk-weighted by the portion economically protected by indirect hyperscaler support (take-or-pay guarantees, minimum-purchase commitments, credit support, strategic dependence). Not a literal dollar subtraction; a risk-weighting framework where hyperscaler support reduces (but does not eliminate) creditor loss probability. Distinguishes nominally-external financing from genuinely-independent financing. Full definition and initial measurement in Ledger Vol 1. **Section I. Stress test and falsifier watch.** Each of the primary stress scenarios and thesis falsifiers updated green/amber/red with observable evidence. **Section J. Transmission matrix update.** Any propagation band revised based on observed propagation during the quarter. The Ledger is designed to be scannable in five minutes and comprehensive in twenty. First volume publishes in early December 2026, covering Q3 2026 with baseline data through Q4 2026 as it becomes available. Every subsequent volume follows the same section structure for comparability across periods. ## 15\. Reconciling with Trillion-Dollar Valuations The $0.65 trillion financing figure sits inside a system whose principal companies carry enterprise values well into the trillions: NVIDIA above $4 trillion, OpenAI at an $852 billion post-money valuation, Anthropic at $965 billion post-money after a $65 billion Series H, memory and networking suppliers in the low hundreds of billions. These quantities measure different things on different bases. ### Three measurement planes | Quantity | What it measures | This essay's scope | | ---------------------------------------- | ----------------------------------------------------------- | ------------------------------------------------------------ | | **Enterprise / equity value** | Discounted PV of all future free cash flows into perpetuity | **Not measured** | | **Corporate capital raised** | Equity + debt at operating-company level | **Partly captured** (via sponsor equity (S5) where relevant) | | **Third-party infrastructure financing** | External financing supporting formation + procurement | **$646B gross** (this essay) | Anthropic being valued at $965 billion does not imply $965 billion of financing has occurred. The $65 billion Series H recapitalises the operating company; only a fraction ultimately reaches third-party infrastructure structures in the current window. The reported Anthropic-Nscale $45 billion six-year compute commitment for the Monarch campus (§6, Example 2) is a much cleaner illustration of the mechanism. The capital waterfall · why $0.65T financing sits inside a trillion-dollar valuation system Five stages of the AI capital system. Each stage measures a narrower subset than the one above. 1\. Enterprise / equity valueStock · discounted PV of all future free cash flows into perpetuityNVIDIA >$4T · OpenAI $852B post-money · Anthropic $965B · Micron \~$140B2\. Corporate capital raisedFlow · equity + debt issued at operating-company level per financing eventOpenAI $110B round · Anthropic $65B Series H · NVIDIA balance-sheet extensions3\. Contractual commitmentsFlow · cloud, compute, prepay, capacity reservations, the mechanism directing capital toward specific infrastructureAnthropic ↔ Nscale $45B / 6-yr / 460MW · Google ↔ Kairos SMR PPA · take-or-pay compute contracts4\. Physical formation + procurementFlow · $0.95T aggregate 2026-28 (this essay), summing physical L1-L3 $0.51T plus compute procurement L4 $0.44TL1a Grid ($73B) · L1b Firm ($13B) · L2 Site ($221B) · L3 Equipment ($204B) · L4 Compute proxy ($437B)5\. Third-party financingFlow · $0.65T gross (this essay) = $0.49T new money + $0.16T refi/secondaryS1 prepay · S2 vendor · S3 GPU debt · S4 RE/PF debt · S5 sponsor equitySTOCKFLOWThese numbers do NOT add vertically.They measure different planes of the same economic system. $965B Anthropic valuation ≠ $65B Series H ≠ $45B Nscale commitment ≠ physical formation ≠ third-party financing. Valuations as of Aug 2026 (Anthropic Series H announcement · OpenAI 2026 round · NVIDIA/Micron public market cap) Read the waterfall top to bottom: enterprise value discounts decades of future cash flows. Corporate capital raised is a small fraction per period. Contractual commitments direct that capital toward specific infrastructure. Physical formation plus compute procurement ($0.95 trillion here) is the aggregate result. Third-party financing ($0.65 trillion here) is the identifiable external subset. **These numbers do not add vertically.** ### Consistency check with L4 NVIDIA data-centre revenue at $400-550 billion in 2026 grows over the window to a three-year total in the $1.3-1.7 trillion range. The L4 procurement proxy filters that through non-hyperscaler share (33% central) and hyperscaler-adjacent overlap (29% central), producing $437 billion central. The trillion-dollar NVIDIA enterprise value discounts a longer horizon of the same revenue. The two quantities are internally consistent; they measure different strips of the cash-flow curve. ### The deeper implication The financing system is detaching from the corporate capitalisation system. Anthropic can be worth $965 billion while raising $65 billion. NVIDIA can be worth several trillion while using its balance sheet to support ecosystem financing. Independent infrastructure developers can raise project debt against leases, contracted compute, and GPUs through structures that sit on no AI-lab balance sheet. That is the financial modularity this essay describes. The Ledger tracks the divergence between market-implied cash-flow expectations and observed financing flows as a leading indicator of demand-side calibration. When enterprise-value multiples expand faster than realised third-party financing capacity, the ecosystem is running ahead of its financing infrastructure. --- ## 16\. Closing The proposition this essay opened with is that AI infrastructure crossed a financing threshold in 2026\. Risk that once sat on hyperscaler balance sheets is now being separated into distinct claims on grid access, physical sites, equipment, compute cash flows, operator equity and model-company demand. Different capital providers underwrite each claim. The bottom-up construction produces three headline measurements and two diagnostic components. Gross third-party financing activity of approximately $0.65 trillion. Third-party-associated formation and procurement of approximately $0.95 trillion. Refinancing and secondary activity of approximately $0.16 trillion. Each is analytically distinct and modelled on its own accounting basis. None are additive to the others. The middle measurement decomposes into physical infrastructure formation L1-L3 of approximately $0.51 trillion and compute and IT procurement proxy L4 of approximately $0.44 trillion; these use different measurement regimes and should be read as parallel diagnostic quantities rather than as a homogeneous formation figure. The architecture that produces these measurements matters more than any of the numbers on its own. The central caveat the Ledger will pursue quarterly: financing separation is not identical to economic risk separation. A hyperscaler that signs a long-term take-or-pay contract may still ultimately bear material downside through minimum-purchase commitments, cancellation penalties, cross-recourse, or strategic dependence even if the nominal financing sits with a neocloud operator. **AI infrastructure financing is becoming contractually modular faster than it is becoming economically independent.** The question for the next phase is who ultimately bears the loss when the underlying AI-demand assumption breaks, rather than who finances the asset. Legal modularity is increasingly observable. Aggregate economic independence is not yet measurable. The Risk Transfer Ratio (H.6) and Recourse-Adjusted Financing (H.7) metrics in the Ledger are designed to measure that distinction over time. Five physical layers each with distinct scarcity constraints. Five primary financing sources each with distinct capital providers, durations, and risk profiles. Insurance capacity and NVIDIA-platform capital as cross-cutting quality modifiers. A transmission matrix that describes how shocks propagate between layers with stock-to-flow conversion into physical formation impact. A cross-layer capital provider concentration pattern that correlates the layers through the capital-provider side even when the underlying assets are independent. A duration mismatch that requires continuous refinancing across technology generations, where the systemic risk sits. The framework is unstable in the sense that the propagation bands change quarter to quarter as new data lands. It is stable in the sense that the layered structure and the transmission mechanism are durable features of the market, not artefacts of the current pricing cycle. The next essay in this series, publishing in September, applies this framework to the Anthropic IPO as a Layer 6 upstream-demand-financing case study. Subsequent essays cover the neocloud sub-category bifurcation (F2), the sponsor-platform layer in depth (F3), the insurance-enhanced credit mechanism (F4), the model-company layer generalised across Anthropic, OpenAI, and xAI (F5), and sovereign and foreign strategic capital as a cross-cutting influence (F6). The recurring Ledger publishes quarterly starting in early December. For now, the answer to how AI infrastructure actually gets financed is this: through separately underwritten claims on physical scarcity, contracted cash flows, and residual value, sitting on a concentrated capital-provider base, with a duration structure that requires continuous refinancing across technology generations. The architecture is what matters. The numbers describe its current state. --- *Series: AI Infrastructure Financing · Volume 1 of 6* *Next in series: F2 · The Neocloud Sub-Category Bifurcates (September)* *Concurrent flagship: The $2 Trillion Anthropic Question (applies the Layer 6 upstream framework, publishing around the S-1 window)* *Recurring column: AI Infrastructure Financing Ledger · Vol 1 · early December 2026* *Every input in the model is tagged with evidence grade, rationale, and falsifier. All numbers regenerate on assumption change. Model and assumptions release with the first quarterly Ledger volume in early December 2026.*